Hull, D., L., Langman, R., E. and Glenn, S., S. (2001) A General Account of Selection: Biology, Immunology and Behavior. Behavioral and Brain Sciences 24 (2): XXX-XXX.

This is the unedited final draft of a BBS target article that has been accepted for publication (Copyright 2000: Cambridge University Press) and is currently being circulated for Open Peer Commentary.

This preprint is for inspection only, to help prospective commentators decide whether or not they wish to prepare a formal commentary.

Please do not prepare a commentary unless you have received a formal invitation indicating that it has been possible to include you in the final list of invited commentators.

For information on becoming a commentator on this or other BBS target articles, write to bbs@soton.ac.uk

For information about subscribing or purchasing offprints of the published version, with commentaries and author's response, write to:

journals_subscriptions@cup.org (North America)
journals_subscriptions@cup.cam.ac.uk (All other countries).

A General Account of Selection: Biology, Immunology and Behavior
 
 

David L. Hull

Rodney E. Langman
Conceptual Immunology Group,
Salk Institute,
La Jolla CA 92037-1099
Electronic mail: langman@salk.edu
URL: http://www.cig.salk.edu
 

Sigrid S. Glenn
Department of Behavior Analysis,
University of North Texas,
Denton TX 76203-0919
Electronic mail: glenn@scs.unt.edu
 
 
 
  David Hull, Dressler Professor in the Humanities at Northwestern University, is the author of five books and over a hundred papers on biological systematics, evolutionary biology, and philosophy of biology. He is past president of the Society for Systematic Zoology, the Philosophy of Science Association, and the International Society for the History, Philosophy and Social Studies of Biology. He is a fellow of the American Academy of Arts and Sciences and the American Association for the Advancement of Science.
 
 
 
 
 
 
  Rodney Langman, Senior Staff Scientist at the Salk Institute for Biological Studies in La Jolla, Californian since 1977, is the author of over 100 papers on experimental and theoretical aspects of immunology, including a monograph entitled 'The Immune System.' He is also a Senior Lecturer at the University of California San Diego.
 
 
 
 
 
 
 
 
 
  Sigrid Glenn Professor of Behavior Analysis at the University of North Texas, is the author over 25 articles and book chapters and 4 books. In the past 10 years, her work has centered on similarities of evolutionary processes in organic, behavioral, and cultural domains. She recently served as president of the Association for Behavior Analysis: International and is founding chair of UNT's Department of Behavior Analysis.

Abstract

Authors frequently refer to gene-based selection in biological evolution, the reaction of the immune system to antigens and operant learning as exemplifying selection processes in the same sense of this term. However, as obvious as this claim may seem on the surface, setting out an account of "selection" that is general enough to incorporate all three of these processes without becoming so general as to be vacuous is far from easy. In this target article, we set out such a general account of selection to see how well it accommodates these very different sorts of selection. The three fundamental elements of this account are replication, variation and environmental interaction. For selection to occur, these three processes must be related in a very specific way. In particular, replication must alternate with environmental interaction so that any changes that occur in replication are passed on differentially because of environmental interaction.

One of the main differences among the three sorts of selection that we investigate concerns the role of organisms. In traditional biological evolution, organisms play a central role with respect to environmental interaction. Although environmental interaction can occur at other levels of the organizational hierarchy, organisms are the primary focus of environmental interaction. In the functioning of the immune system, organisms function as containers. The interactions that result in selection of antibodies during a lifetime are between entities (antibodies and antigens) contained within the organism. Resulting changes in the immune system of one organism are not passed on to later organisms. Nor are changes in operant behavior resulting from behavioral selection passed on to later organisms. But operant behavior is not contained in the organism because most of the interactions that lead to differential replication include parts of the world outside the organism. Changes in the organism's nervous system are the effects of those interactions.

The role of genes also varies in these three systems. Biological evolution is gene-based (i.e., genes are the primary replicators). Genes play very different roles in operant behavior and the immune system. However, in all three systems, iteration is central. All three selection processes are also incredibly wasteful and inefficient. They can generate complexity and novelty primarily because they are so wasteful and inefficient.
 
 

Keywords:

Evolution, immunology, interaction, operant behavior, operant learning, replication, selection, variation.
 

1. Introduction
 

What was so radical about Darwin's theory of evolution? In the following quotation, B. F. Skinner (1974, p. 40-41; See also BBS special issue ) dismisses the usual answers given to this question and suggests a nonstandard answer of his own:

Darwin's theory of natural selection came very late in the history of thought. Was it delayed because it opposed revealed truth, because it was an entirely new subject in the history of science, because it was characteristic only of living things, or because it dealt with purpose and final causes without postulating an act of creation? I think not. Darwin discovered the role of selection, a kind of causality very different from the push-pull mechanisms of science up to that time.

Although we think that the late appearance of selection on the intellectual scene no doubt had numerous causes, we agree with Skinner that part of the answer is surely the counter-intuitive kind of causality exhibited in selection processes. Push-pull causation does seem "natural" to us. So does functional organization. But the action of selection processes does not. This fact about how people in the West think is reflected in natural languages. Finding terms to describe selection processes that do not have all sorts of inappropriate connotations is not easy.

Numerous biologists and philosophers of biology have presented analyses of gene-based selection in biological evolution (e.g., Lewontin 1970; Dawkins 1976; Hull 1980; Sober 1984; Vrba & Gould 1986; Lloyd 1988; Sober & Wilson 1998), but relatively few workers have tried to present a general account of selection to see which processes in addition to gene-based biological evolution are genuine selection processes and which are not. (The chief exception is Darden & Cain 1989). Are selection processes sufficiently different from other sorts of causal processes to warrant a separate analysis? The sort of selection that goes on in biological evolution is surely an instance of selection, but how about other putative examples of selection, e.g., the reaction of the immune system to antigens, operant learning, the development of the central nervous system, and even conceptual change itself (Cziko 1995)?

In this target article we provide a general account of selection. The chief danger of such "general analyses" is that they can be either too broad or too narrow. If the account is too broad, then everything becomes a selection process-including crystal formation and balls rolling down inclined planes. We have no objection to anyone attempting to present general accounts of more global phenomena, such as the persistence of patterns, but we limit ourselves just to selection processes. The other danger is to make the analysis too narrow so that each putative type of selection becomes unique. For example, the genotype/phenotype distinction plays a central role in gene-based selection in biology. Is this role common to all selection processes or unique to selection at the biological level? Being able to distinguish self from nonself is crucial in the immune system. Is the self-nonself distinction also important in other sorts of selection? In operant learning selection occurs only with respect to sequences of environmental interaction rather than with respect to numerous concurrent alternatives. Is this difference sufficient to disqualify it as a case of selection?

Such questions cannot be answered a priori. We have to try various analyses and see how they turn out. Our goal is to see if selection processes can be construed usefully as a special sort of causal process. The success of such an analysis will be determined by the use that those scientists working on various sorts of selection can make of it. If they find that our analysis helps them to understand the sort of selection they are studying more clearly, then it has succeeded; if not, then it has failed (for a defense of the method of abstraction, see Darden & Cain 1989). Even though causation is absolutely central to our understanding of selection, we do not attempt to present a general analysis of causation in this paper. In the past, some of the disputes that have arisen with respect to selection actually turn on different views of causation (e.g., Sober 1984, 1992; Brandon 1982, 1990; Brandon et al. 1994; van der Steen 1996; Glymour 1999). Ideally, we should include an analysis of "causation" alongside "selection." However, every analysis must stop somewhere. Not all of the substantive terms used in an analysis can themselves be analyzed. We do not present an analysis of causation in this paper because the literature is too vast and the alternatives too various. For better or for worse, in this paper we depend on the reader's largely tacit understanding of this extremely basic notion. The most that we can do in the space of a single paper is to point out when different notions of causation have caused problems, as in the instance cited above (for a recent discussion of causation, see Salmon 1998).

The three authors of this target article come from three very different backgrounds. David Hull (1980, 1987) emphasizes his work on gene-based selection in biological evolution, treating selection as an alternation between replication and environmental interaction. Sigrid Glenn contributes her work on operant learning as a selection process (Glenn 1991; Glenn & Field 1994; Glenn & Madden 1995). Rod Langman (Langman 1989, Langman & Cohn 1996) adds his extensive theoretical analysis of the immune system. In this paper we strive to pool our conceptual resources to produce a general account of selection adequate for the three sorts of selection under investigation-gene-based selection in biological evolution, the reaction of the immune system to antigens, and operant learning.

We do not offer an analysis of three other possible examples of selection--the development of the central nervous system, social learning, and conceptual change. We do not include an extensive discussion of neuronal development because the empirical facts remain too controversial (Edelman 1987; Quartz & Sejowski 1997). Once neurophysiologists have worked out the basic structure of neuronal development, we then will be in a position to evaluate this process to see if it can legitimately count as a selection process. If it fits our analysis, well and good. If not, then either neuronal development is not a selection process or else our analysis is deficient. The second example of a putative selection process that we do not discuss in this paper is social learning, even though social learning is one of the most commonly cited examples of a selection process. Instead we limit ourselves to individual learning as a selection process. The strategy that we have adopted is to deal with the simplest cases first. Once we understand the most unproblematic instances of selection, we can then turn to the more difficult cases. The same justification applies to conceptual change, including conceptual change in science. Is the process that has allowed the three of us to understand selection in biological evolution, immunological reactions, and operant learning itself a selection process? Is conceptual change itself a selection process? Although we find the construal of conceptual change as a selection process fascinating, we do not discuss it in this paper (see Laland, Olding-Smee & Feldman forthcoming).

From the start, we have to register one warning: none of us claims to present the standard interpretation of the processes with which we are dealing, mainly because no such "standard" interpretation exists in any of the three cases that we investigate. For example, numerous objections have been raised to neo-Darwinian versions of evolutionary theory, especially the heavy emphasis that is placed on genes and the cavalier attitude frequently exhibited toward the environment. With a few noteworthy exceptions (e.g., Brandon 1990), the environment is treated as an unarticulated background against which selection operates. With respect to the immune system, considerable disagreement exists concerning the mechanism that allows the immune system to react selectively against nonself but not self components (e.g., Silverstein & Rose 1997). Numerous versions of learning theory can be found in psychology. Even if one limits oneself just to operant learning, disagreements exist. Can the stimulus that functions with respect to behavior be inside as well as outside the organism?

As much as scientists strive to reduce the amount of disagreement in science, they never come close to succeeding, and if science itself is a selection process, they cannot. In this paper we could not examine all versions of all of the theories that we treat. We had to select one from each of the domains. The issue is whether this version and others like it can be properly construed as selection processes, not whether we accept the reader's preferred version. This caveat applies with special force to theories of operant behavior. Some psychologists reject such theories out of hand. Others have strong preferences for one version over all others. In this paper we cannot answer the objections that have been raised to any of the three broad range of theories we discuss. Instead, for the purposes of this paper, we accept their overall adequacy and proceed from there to decide whether or not they exemplify a particular sort of process-selection. The point of this paper is not the choice of the one and only correct version of any of the theories that we treat. It is to discover if theories of this type can be construed as exemplifying selection processes.

Yet another problem that we confronted in writing this collaborative paper is that the three of us used very different terms to describe what we took to be the same sort of process. From the outset, we had to reduce differences that were mainly terminological-a task that turned out to be much more difficult than we had anticipated and not fully completed even now. One danger was allowing the process of biological evolution to play too large of a role in our undertaking. Because selection processes were first worked out in gene-based biological evolution, the temptation is to take it as standard and compare other candidates to it, but such a strategy would be biased. Historical precedence does not guarantee conceptual priority. In this paper we need to investigate each candidate in its own right, rather than taking gene-based selection in biological evolution as the standard by which all other putative examples of selection processes are to be evaluated. Even the use of the phrase "selection in gene-based biological evolution" is misleading. Both the functioning of the immune system and operant behavior are to some extent "gene based" and "biological." However, they also include processes that are not "gene based" in this narrow sense. But for want of a better name, we retain the phrase "gene-based selection in biology."
 

2. A Brief Characterization of Selection
 

Several authors have attempted to characterize selection in as brief a fashion as possible. For example, Campbell (1974) describes selection as a function of blind variation and selective retention, while Plotkin (1994, p. 84) characterizes it as a matter of generation, testing and regeneration. The trouble with these characterizations is that they are too brief. If one wants to understand selection, a sentence or two, no matter how succinct, will not do. Understanding space and time requires more than looking up these terms in a dictionary or in a physics text. Instead one must learn the relevant physics. Similarly, anyone who wants a deep understanding of selection has to study this phenomenon. Just inspection of a brief characterization of the process will not do. This much being said, we define selection as repeated cycles of replication, variation and environmental interaction so structured that environmental interaction causes replication to be differential. The net effect is the evolution of the lineages produced by this process. Each word in this definition needs careful explication. The message is not to be found in the preceding brief characterization of selection but in the ensuing discussion.
 

2.1. Variation

Variation is sometimes considered part of the selection process (Darden & Cain 1989), sometimes as a precondition for selection processes (Hull 1980). Either way, variation is absolutely essential for the operation of selection processes. If there is no variation, then there are no alternatives to select among. However, the characterization of the variation that functions in selection processes has been one of the most contentious topics in the literature-and the most frustrating. It seems that no adjective exists in the English language that accurately reflects the sort of variation that occurs in selection processes. Is this variation blind, chance, random, nonprescient, nondirected, nonteleological, unforesighted, what?

First and foremost, the variations that function in selection processes of all sorts are caused-totally caused. No one writing in this literature feels inclined to introduce miracles in their descriptions of variation. The task is to describe the sorts of causes that produce this variation. When advocates of selection say that the variations that are operative in selection are "blind," they cannot possibly be using this term in a literal sense, as if some variations can see and others cannot. They must mean it in some metaphorical sense. When they term variations "chance" or "random," they cannot be using these terms as they are defined in mathematics. The requirements specified in these definitions are so rigorous that few, if any, natural phenomena can meet them.

Evolutionary biologists are well aware of the various factors that cause mutations. They are also aware that these mutations frequently depart from anything that might be termed "pure randomness." In fact, in many cases the very biologists who insist that the variations that function in selection processes are "random" are the ones who discovered these departures from randomness in the first place. For example, mutations that produce melanic forms crop up in certain groups of organisms with a greater frequency than the laws of chance would allow. On certain chromosomes hot spots exist that exhibit extremely high rates of mutation. For example, whole segments of immunoglobulin genes have bursts of mutation 10⁶ fold greater than average (see Dawkins 1996, p. 80-82 and Pennisi 1998 for additional examples). The other adjectives used to modify "variation" arise in the context of selection in conceptual change, but no one thinks that people, including scientists, are "prescient." People may try to anticipate the future, we can even predict the future in some cases, but no one is literally prescient.

Confusion in these matters stems in large part from the legacy of the early days of evolutionary biology, in particular the controversy between the Darwinians and Lamarckians. Critics of Darwinian versions of evolutionary theory tend to term any departures from the simplest forms of inheritance "Lamarckian." To be sure, numerous forms of nonstandard inheritance have been discovered over the years (Crow 1999). The issue is whether or not any of these fascinating forms of inheritance are in any significant sense "Lamarckian." The distinction between Darwinian and Lamarckian inheritance depends on the distinction between genotype and phenotype. According to the inheritance of acquired characteristics, the environment modifies the phenotype of an organism so that it is better adapted to the environmental factors that produced this phenotypic change in the first place-better adapted than those organisms that were not modified in this way. This phenotypic change is then transmitted somehow to the genetic material so that it is passed on in reproduction. Thus, according to this view, species can rapidly adapt to environmental change. In Darwinian evolution inherited variations are "random" with respect to (i.e., independent of) the effects that they produce, while in Lamarckian evolution they are not.

Both aspects of the preceding discussion need emphasizing. First, in Lamarckian evolution, the phenotypic change that results must make the organism better able to cope with the environmental factor that produced the phenotypic change in the first place. They must be adaptations. Exposing the skin to increased sunlight causes it to darken so that the organism is better able to withstand increased sunlight. Second, in Lamarckian evolution the phenotypic change must be transmitted to the hereditary material so that it can be passed on genetically. A mother dog giving fleas to her puppies is not an instance of the inheritance of acquired characteristics because it is not an instance of inheritance in the sense required by Lamarckian inheritance. Biologists do not have a corner on the term "inheritance." Other workers can and do use it in a variety of other senses. Our discussion, however, concerns Lamarckian inheritance as a biological phenomenon (for a sampling of the recent literature on Lamarckian forms of inheritance, see Lenski & Mittler 1993; Jablonka & Lamb 1995; Rosenberg, Harris & Torkelson 1995; MacPhee & Ambrode 1996; Peck & Eyre-Walker 1998; Benson 1997; Andersson, Slechta, & Roth 1998).

In sum, statements about the sorts of variation that function in selection processes need not include any reference to their being "blind," "random," or what have you. All of the terms that have been used to modify "variation" are extremely misleading. Hence, we see no reason to put any adjective before "variation" in our definition of selection. Our analysis concerns only those instances in which variations occur, without regard to their eventual contributions to fitness in biological evolution or some corresponding circumlocution with respect to the immune system and operant behavior. In this paper we deal with natural selection as it functions in Darwinian evolution today. Darwin himself included Lamarckian forms of inheritance in his theory, but Darwinians today do not. Darwinian evolution is currently limited to Darwinian (or Weismannian) inheritance. If Lamarckian forms of inheritance turn out to exist, we have no doubt that these mechanisms will be promptly incorporated into the Darwinian theory the way that neutral mutations were.
 

2.2. Replication

Replication is the second important notion in our brief characterization of selection, and it poses as broad a spectrum of problems as does variation. Replication contains two elements-iteration (or repetition or recursion, depending on one's terminological preferences) and information. Early on, Dawkins (1976) published a highly influential general account of selection that emphasized the role of "replicators." They are the entities whose structure contains the "information" that is passed on differentially in selection. The structure of replicators counts as information in the sense that it codes for the character of the individuals (or "vehicles") that the replicators produce. The only variations in the structure of replicators that matter are those that modify the relevant vehicles. These vehicles then interact with one or more local environmental conditions. Some of these variants survive to replicate and the process begins again. That is why Plotkin (1994, p. 84) in his analysis of selection emphasizes generation and regeneration. However, sequential replication is not enough. Variants must be linked to proliferation so that at any one time, numerous alternatives are available for selection. At the very least, the frequency of replicators must change sequentially through time.

The only feature of the analysis of selection-type theories provided by Darden and Cain (1989, p. 110) with which we disagree is the demotion of iteration to an ancillary feature of selection. For them, selection is essentially a one-shot deal that can be, but need not be, repeated. They replace iteration with such evaluative notions as benefitting and suffering: "Several types of effects result from the differential interactions. In the short range, individuals benefit and suffer." Although they realize that such terms as "benefit" and "suffer" sound anthropomorphic and value-laden, they have to introduce them because they do not treat iteration as central to selection. A single cycle of replication and environmental interaction would fulfill the requirements of their analysis, just so long as it hurt or helped the relevant individuals.

In our analysis we are able to avoid the use of such problematic notions as "benefit" because of the central role of iteration. If some characteristic is increasing in frequency, then it is very likely (though not necessarily) doing some good. It is better adapted to its environment than other variants. According to our account, Darden and Cain's single-cycle analysis of selection is (at most) a limiting case of our account (see section 4.3 under "big-bang" for further discussion). One reason why we prefer no mention of benefit and harm in our general account of selection is that their elimination from explanations of biological adaptations was one of Darwin's major achievements. We are not inclined to reintroduce such notions at this late date if we can avoid it. Iteration has problems of its own (e.g., how to keep "survival of the fittest" from degenerating into a tautology), but these problems can be handled with only a modicum of care and effort (see Lipton & Thompson 1988).

Replication is inherently a copying process. Successive variations must in some sense be "retained" and then "passed on." In many earlier definitions of "selection," all that is required is heritability, not genealogical inheritance. As Thompson (1994, p. 638) observes with respect to gene-based selection in biological evolution, natural selection "does not require genes or even direct descendants; all it requires is that the presence of a configuration of elements in one generation makes more likely the presence of the same configuration in the next generation." We agree with Thompson as far as genes are concerned but draw the line at descent. In biological evolution, replication is accomplished by molecules of DNA splitting and the missing nucleotides being filled in so that the information contained in the resulting molecules is retained. This is one way for replication to occur, but it is only one way. If splitting and reassembly is considered to be essential to all selection processes, then only gene-based selection in biological evolution and the functioning of the immune system count as selection processes. We think that this restriction it too narrow. A variety of mechanisms exist that can have the same effect as splitting and reassembly.

We have taken the opposite tack with respect to descent. Mechanisms other than modification through descent could serve the function that descent does. However, thus far, descent is the only mechanism that has evolved to produce the correlations necessary for selection. A more general analysis than ours might be couched in terms of retention of pattern or configuration from one generation to the next. However, in the absence of replication, the notion of "generation" becomes extremely problematic. In our analysis, we emphasize the mechanisms that produce evolutionary change, not just correlations. The preceding discussion is just one instance of the problems that arise in conceptual analysis. Is our analysis too broad or too narrow? Others might well make decisions different from ours, decisions that might have considerable merit.

In our analysis, the first component of replication is iteration (or repetition or recursion). The second is information. As Williams (1992, p. 11) points out, structure is necessary for selection, but structure alone is not good enough. Some of this structure must count as information. With respect to gene-based selection in biological evolution, "A gene is not a DNA molecule; it is the transcribable information coded in the molecule." DNA exhibits numerous structural elements. For example, it forms a double helix, and the bonds that connect the two bases that make up each of the rungs of the DNA ladder are easier to sever than those that connect successive nucleotides. With respect to gene-based selection in biological evolution, the preceding features of the DNA molecules count as structure but not as information. Of course, DNA itself had to evolve via selection. DNA molecules are adapted to replicate. The features of DNA molecules that allow them to replicate were selected for in the origin of life (Küppers 1990). But these features of DNA molecules do not "code for" anything.

In gene-based selection in biological evolution, much of the relevant information is comprised of the linear sequence of bases in molecules of DNA. Unfortunately, in spite of the massive amount of work done by a variety of scholars on explicating the notion of information, none of the suggestions made thus far is adequate to distinguish information as it functions in selection processes from other sorts of structure. For example, physicists treat any structure as "information." The information contained in a double helix is no different in kind from that exhibited in the linear sequence of bases. As helpful as the work of Dretske (1981) and Küppers (1990) may be in other respects, it cannot be used to distinguish the special sort of structure exhibited by sequences of base pairs in molecules of DNA from structure as such. Nor is it adequate to make this crucial distinction with respect to the immune system and learning. The one bright spot on the horizon is that several biologists, such as John Maynard Smith (2000), and philosophers of biology such as Peter Godfrey-Smith are currently working on the problem. Progress may be forthcoming. If we are to have an adequate conception of selection, progress in our understanding of information must be forthcoming. In the case of causation, the problem is that too many different analyses of causation exist, some adequate for certain causal situations, others adequate for others. In the case of information, the problem is that too few analyses of information exist, and none of them is adequate for understanding selection processes. In writing this paper, we were presented with two choices: register this major deficiency in our understanding of selection and move on or present from scratch an analysis of information that is up to the task. We decided on the first alternative. We hope that others will eventually come to adopt the second alternative (for a critical evaluation of the recent literature on information theory, see Sarkar 1996 and Harms 1998).
 

2.3. Environmental Interaction

Dawkins (1976) placed considerable importance on the notion of replication. It is the primary explanatory concept in his analysis of selection. Many critics think that Dawkins places too much emphasis on replication as if it were sufficient for selection. They also raise the issue of the problematic character of information, as we have. Dawkins also introduced a second notion, that of a "vehicle." According to Dawkins, replicators replicate themselves (homocatalysis). In addition, they produce vehicles (heterocatalysis). Replicators do more than just cause or produce vehicles; they "code" for them. For Dawkins the relation between replicators and vehicles is that of development. A third major criticism of Dawkins's view of biological evolution turns on the relation that he sets out between replicators and vehicles. Replicators not only code for their vehicles but also ride around in and steer them. Vehicles are nothing but survival machines, lumbering robots controlled by the replicators that produced them.

To begin with, Dawkins's vehicles of selection have to be distinguished from Campbell's (1979) physical vehicles. For Campbell "vehicle" refers to the material basis or carrier of information; e.g., molecules of DNA that incorporate information in the order of base-pairs, the paper on which books are printed, the plastic that was once used for phonograph records, and the chips in electronic computers. Clearly, Dawkins means something else by "vehicle." Most narrowly, he means the organisms produced by genomes. Needless to say, this narrow notion immediately raises the nature-nurture issue. In what sense does a genome code for an organism? A genome all by itself never produced anything (Marx 1995). Genomes plus numerous other factors produce organisms. However, according to the standard framework, both genes and environmental conditions cause traits, but only genes code for them. Of course, the metaphor of genes "coding for" traits remains as problematic as ever. We are well aware that sketches of several alternatives to the traditional gene-based view of biological evolution exist. Our concern in this paper is to provide an account of selection adequate for the traditional view, not to answer every objection raised to the traditional view.

If the general analysis presented in this paper is to be applied to specific instances, the terms used in this analysis must be "operationalized." Environmental interaction must cause replication to be differential. For example, drift is differential perpetuation without environmental interaction. In this connection, "selection for" is often distinguished from "selection of." A gene contributes to the development of a trait that interacts with the organism's environment so that this gene replicates more profusely than the genes of conspecifics that lack this gene and trait. This gene is being "selected for" this ability. A second gene adjacent to the first gene may piggyback on it. Because this second gene does not interact in the relevant sense with its environment, it is not part of the cause of this increase in frequency. As we construe selection, development is only one of the causal relations that can exist between what Dawkins terms "replicators" and "vehicles." For a truly general account of selection, a much broader relation is necessary. The relation must be causal, but it need not be developmental. Numerous other processes are also operative. For example, molecules of DNA interact with their environments to replicate themselves, but this process does not involve anything like ontogenetic development in the production of vehicles.

If the distinction between replication and environmental interaction does anything, it goes a long way in resolving the levels of selection controversy. When Dawkins (1994) says that genes are the units of selection, he means replication. Genes are the primary units of replication and "hence" selection. When others such as Mayr (1997) say that organisms are the primary focus of selection, they mean environmental interaction. In gene-based biological evolution, organisms are the primary units of environmental interaction and "hence" selection. To be sure, both replication and environmental interaction are necessary for selection, but we do not think that either is sufficient by itself. Both are needed for selection to occur. As Lloyd (1988) has pointed out, the levels of selection controversy concerns environmental interaction, not replication. Entities from molecules of DNA, cells and organisms to colonies, demes and possibly entire species interact with ever more inclusive environments in ways that bias replication. Selection involves two processes and not one. There are units of replication and units of environmental interaction, but there are units of selection only in a highly derived sense, in the same derived sense that IQ is a measure of intelligence (Hull 1980; Heschel 1994).
 

3. Selection in Biological Evolution
 

The highly general characterization of selection set out in the preceding pages applies in a straightforward way to selection in gene-based biological evolution. In each case the sort of selection that population biologists study can be seen to be a special case of the more general analysis of selection provided in this paper (for a recent criticism of analyzing selection in terms of replication, see Griesemer 1999).
 

3.1. Mutation and Recombination

In gene-based biological evolution, the sources of variation are point mutations and recombination. Point mutations result in a single nucleotide being changed. Recombination results from the reorganization of the linear structure of DNA. As it happens, recombination produces most of the variation that is actually operative in biological evolution. The linear sequence of nucleotides in DNA provides the information necessary for the production of proteins. Any rearrangement of these orderly nucleotide sequences stands a chance of changing the genetic information encoded in its DNA and possibly the phenotype of the organism as well. The causes of variation in the genetic material are important. The effects that genes have on the phenotype of an organism are equally important. In selection, genetic variations must result not only in phenotypic variations, but also these differences must affect the individual with respect to survival and/or reproduction.

At one time biologists believed that the vast majority of mutations result in a decrease in proliferation, while only a small percentage increase proliferation or do not affect it at all (but see Peck & Eyre-Walker 1998). Mutations can fail to affect proliferation in two ways: either they have no phenotypic effects or else the phenotypic effects make no difference to survival and/or reproduction. Once biologists had more direct access to the genetic material, they discovered all sorts of unexpected things about it. They found that most of the genetic material has no apparent function. Perhaps it did in the past, perhaps its current functions have yet to be discovered, but right now most of the genetic material does not seem to do much of anything. In part as a result of the former finding, it turns out that most mutations are selectively neutral (i.e., as a result of environmental interaction, they neither increase nor decrease in frequency), while some are selected against (they decrease in frequency because of environmental interaction), and only a small percentage are actually selected for (they increase in frequency because of environmental interaction). As important as Kimura's (1983) work has proven to be, his claim that changes in our beliefs about the relative frequencies of these three types of mutation requires a "new" theory of evolution has not been widely accepted (see Brookfield 1995).
 

3.2. Replication

What are the primary replicators in biological evolution? Genes, larger chunks of the genetic material, and sometimes even entire chromosomes can function in replication. Replication at higher levels of organization may also occur, but the more inclusive the entity, the harder it is for the requirements of replication to be met. The important point is that once the notion of replication has been distinguished clearly from environmental interaction and selection, this question (and it is an empirical question) can be answered more definitively. What are the entities that interact with the environment in ways that result in differential replication? Everything from genes, cells and organisms to hives, demes and possibly entire species. Environmental interaction wanders up and down the organizational hierarchy, while replication is largely limited to the genetic material. In some circles the view that genes are the primary replicators and that environmental interaction occurs at a variety of levels is considered radical-possibly true but still in need of extensive elaboration and corroboration. In other circles, it is considered to be the received view that needs to be replaced by a more sophisticated theory.

Needless to say, wide agreement does not exist about the character of this more sophisticated view. As is usually the case in such disputes, one side parodies the other. For example, certain critics of the received view treat replication as a nonsense notion, as if replication is supposed to occur in the absence of any and all environmental contributions, but even the most rabid gene replicationist knows all of this. Quite obviously, replication requires all sorts of environmental inputs, including energy and the relevant enzymes (Marx 1995). Traditional versions of neo-Darwinian theory have enough faults without inventing irrelevant parodies. Perhaps evolutionary biologists have not spent enough time attempting to integrate development into evolutionary theory, but they are well aware of its existence and the need for such an integration (e.g., Davidson, Peterson & Cameron 1995). Perhaps an adequate theory of evolution will require the sort of fundamental revisions that some critics of the received view suggest (e.g., Jablonka & Lamb 1995; and Griesemer 1998), but evolutionary biologists are likely to be swayed more by positive contributions than by continued criticism.
 

3.3. Environmental Interaction

In the traditional view of biological evolution, the primary means of recording (or retaining) and passing on variation is via genes. That is why we have been terming selection in biological evolution "gene-based." Then these genetic variants must interact either directly or indirectly with the environment so that in the last analysis replication is differential. Some replicates are more likely to be passed on than others. In addition to replicating, genes also code for phenotypes, and these phenotypes can be exhibited at various levels in the organizational hierarchy from genes, cells and organisms to colonies, populations and possibly entire species. Genes interact with their cellular environments, but they also interact with increasingly more complex environments via their surrogates. The "fit" between these phenotypes and their environments determines which genes get passed on and which not. In more general terms, the information contained in replicators gets passed on differentially because of how successfully they or their products interact with their respective environments (Brandon 1982).

What are the entities that function in environmental interaction? Can we get along just with the notion of phenotypic effect, regardless of these effects being bundled together into organisms? As strange as it might sound, genes themselves exhibit adaptations. The most obvious thing about DNA is that it is adapted to replicate. During periods of replication, genes interact with their immediate environments. They could not replicate without appropriate environmental contributions. Organisms exhibit phenotypic traits in the most obvious sense. Some organisms in a species have split telsa; others do not. At the other extreme, even species exhibit phenotypic traits. For example, the peripheries of the ranges of some species are highly convoluted. If speciation usually occurs at the peripheries of these ranges, then such convolutions, if they are heritable, might count as adaptations for increased rates of speciation. Some authors complain that requiring adaptations for selection, including species selection, is too restrictive. A more general notion is required, the sort of general characterization that we have provided (Lloyd 1988; Wilson & Sober 1994; Sober & Wilson 1998; and Gould & Lloyd forthcoming).

Much of the discussion of selection in the recent literature has concerned replication, but environmental interaction is at least as important in selection as is replication. The strongest feature of Darden and Cain's (1989) analysis of selection is the emphasis that they place on environmental interaction. As they put it, "individuals must be in an environment with critical factors that provide a context for the ensuing interaction" (Darden & Cain 1989, p. 110). The debate that Dawkins's The Selfish Gene (1976) elicited was generated in large measure by an ambiguous use of the term "selection" in the literature. One side of the dispute conflates "replication" with "selection," while the other side conflates "interaction" with "selection." Dawkins argues at great length that in biological evolution the relevant replicators are genes and only genes. Replication is certainly necessary for selection as it occurs in biological contexts, but it is not sufficient. Replication and variation in the absence of environmental interaction results in drift, and as important as drift may be in the evolutionary process, it is not a consequence of selection (Donoghue 1990). Selection requires an interaction of some sort between the environment and the replicating entity.

Dawkins's opponents have countered that organisms are the primary focus of selection. They, not genes, are the units of selection. Just as Dawkins, early on in the controversy, too often elided from replication to selection, his critics tended to equate selection with environmental interaction. As in the case of replication, environmental interaction is necessary but not sufficient for selection. Without replication, iteration is impossible, and in the absence of iteration, selection could not be cumulative. Selection is the result of differential replication caused by environmental interaction. Once again, selection is two processes, not one. It is the alternation of replication and interaction with the occasional introduction of variation.
 

3.4. The Environment

Of all the terms in the preceding characterization of selection, "replication" has received the greatest attention. However, the most difficult notion is that of the environment. Not until Antonovics, Ellstrand and Brandon (1988) has it received the analytic attention that it deserves (see also Brandon 1990). These authors distinguish between three different sorts of environment-the external, ecological, and selective environments. The external environment is the "sum total of the factors, both biotic and physical, external to the organism that influence its survival and reproduction" (Brandon 1990, p. 47). The ecological environment of an organism is composed of "those features of the external environment that affect the organism's contributions to population growth." (Brandon 1990, p. 49). Finally, the selective environment is an area (or population) that is "homogenous with respect to the relative fitness of a set of competing types" (Brandon 1990, p. 69).

One problem with respect to selective environments is whether or not to include other organisms, including conspecifics, as part of the selective environment. Such decisions have effects, for example, on how one handles cases of density dependent population regulation (Brandon 1990, p. 65). Wilson and Sober (1994, p. 641) see this issue as clearly distinguishing their views from those of Dawkins. No sooner did Dawkins introduce the notion of a "vehicle" in his account of selection than he began to undermine it. 'I coined the term "vehicle" not to praise it but to bury it' (Dawkins 1994, p. 617). According to Dawkins (1994, p. 617), "Natural selection favors replicators that prosper in their environment. The environment of a replicator includes the outside world, but it also includes most importantly, other replicators, other genes in the same organism and in different organisms, and their phenotypic products." Wilson and Sober (1994, p. 641) respond that Dawkins' goal of reconceptualizing vehicles of selection as part of the external environment (in Brandon's sense) reveals a deep contradiction in the gene-centered view of selection. Clearly the notion of environmental interaction deserves at least as much attention as replication.
 

4. Somatic Selection in the Immune System
 

More than a million different antibodies are needed to provide sufficient protection against the huge number of pathogens a host may encounter during its lifetime. Antibodies are protective because they act as markers that signal the recruitment of powerful bio-destructive cells and enzymes which then destroy the pathogen and stop it from over-growing the host. Managing to make sure that none of the millions of antibodies recognize any part of the host is obviously essential. If there were only a few antibody specificities, and a correspondingly small number of genes encoding these antibodies, then any rare cases of "self" recognition might reasonably result in the destruction of that rare organism-this is an example of germline selection. When the rate of evolution of the pathogen (often hours) is much faster than the rate of evolution of the host (often months to decades), then the host genome cannot carry the millions of different genes needed to track the millions of different mutations in the pathogens. Moreover, among the millions of different antibodies, some will inevitably recognize a self-component of the host and have the potential to destroy the host. What makes the immune system special is that it is able to select on the specificity of each antibody and eliminate the deleterious anti-self before it can actually kill the host. Because each different antibody is expressed in a different cell with a correspondingly different set of genes that encode that antibody, the immune system is able to select on the cell in order to eliminate these anti-self antibodies instead of having to eliminate the whole organism. This form of cellular selection on genetic variants is an example of somatic selection. In immunology it is common to refer to the germline as the genetic material that is selected upon when individuals are replicated and to distinguish this from the soma where the genetic material of individual cells can be varied and selected upon as cells are replicated. While the factual basis for phenomena discussed here can be found in any modern textbook of immunology, the conceptual analysis, should not, however, be taken as representing the standard view of the immune system.
 

4.1. Variation: The origins of antibody diversity

The genetic basis of antibody diversity is partly due to the presence of several different, normally inherited genes, and partly due to mutations that occur in these genes when they are expressed in the soma as antibody-producing B cells. Extensive genetic and sequencing studies can be summarized along the following lines. The antibody molecule is made up of two different polypeptides, the L (light) and H (heavy) chains, which are encoded at two different genetic loci. The particular specificity of an antibody is determined by roughly equal contributions from the L and H chains. The part of each chain that is primarily concerned with antibody specificity is called the V (variable) region and the remainder the C (constant) region. Each region is encoded as a separate gene segment, and there are about 100 V-L and 100 V-H gene segments but only one C-L and one C-H segment. A series of gene fusions permanently changes the chromosomes in B cells and results in the joining of any one of the 100 V segments with the single C segment to produce a single V-C gene that encodes the complete L or H polypeptide. The 100 different L chains and 100 different H chains form random pairs and 10,000 corresponding different specificities. The gene fusions are arranged in such a way that joining errors are maximized. Consequently, few B cells are actually able to produce two L or two H chains. In other words, the B cell is made functionally haploid so that each B cell expresses only one kind of LH pair and, therefore, one specificity. Of course the level of waste is relatively high as 70 to 90% of B cells that attempt to produce antibodies fail and are eliminated.

Throughout the life of an organism, the B cell population is undergoing constant renewal, and this renewal requires the mechanism for eliminating potentially self-reactive B cells to operate continuously throughout life. Controversy surrounds the details of this mechanism of self-nonself discrimination, but the exact nature of this mechanism is unimportant here. The result in any case has to be that individual B cells can be somatically selected according to the particular antigens that react with their receptors. The result is a means of selecting against B cells that can react with self components and neutral selection on B cells with specificities that do not react with self components. However, when a B cell that has not reacted with self is subsequently confronted by the particular pathogen with which it can react, then the B cell is strongly selected for, and, so long as antigen persists, the cells undergo many rounds of mutation and division while secreting huge amounts of their antibodies. The negative selection pressure imposed by self components is constant (self is constant). During the many rounds of cell division that occur when a B cell is under selection by nonself antigens, mutations are introduced in the V segments of the L and H genes that, by chance, affect specificity. These mutations are so important that a special mechanism operates over the V gene segment and is able to introduce single base changes at the rate of 10^-3 per base pair per generation; in contrast, the normal rate of mutation of around 10^-9 per base pair per generation operates on the C segments. Some mutations in V segments are neutral and do not affect specificity, others destroy function, and a few change specificity and improve the ability of the antibody to react with its antigen at much lower concentrations than were present when the B cell was initially selected.

A brief comment on some terms and concepts might be helpful. Antigens are the parts of the pathogen that react with antibodies. Usually the term "paratope" is used to describe the part of the antibody that binds the antigen, and the term "epitope" is used to describe the site on the antigen that reacts with the paratopic part of the antibody. A complex pathogen, such as a bacterium, can expose many epitopes and induce the production of many paratopes, including mutant forms of the initially selected paratopes. The actual B cells that are selected by a particular antigen will depend on the concentration of the antigen and the affinity of the B cell receptor for that antigen. As a result, some B cells will respond only at high antigen concentrations while others will respond only at low antigen concentrations.
 

4.2. The Replication-Variation-Interaction Sequence

The course of events following infection by a pathogen begins with a small inoculum of dividing pathogens. Initially they are at too low a concentration to cause the selection of any B cells. Then, after some time, the numbers of pathogens increases to reach a concentration that can induce an antibody response. The responding B cells proceed to divide and secrete huge amounts of antibody. As the antibody diffuses into the body fluids, it binds to the pathogen and so marks it for destruction. Providing enough antibody is present to halt the growth of the pathogen, then the immune system will have protected the host. As the numbers of pathogen decrease, the concentration of antigen driving the division and mutation of B cells also decreases with the net result that only those B cells able to respond to the lowest concentrations of antigen will remain dividing. In this case somatic selection for variant antibody genes allows some B cells to divide more often than others. During this period, B cells will have behaved in a manner very similar to the pathogenic organisms. Each will have also undergone mutation, and those pathogens that could render their antigens unrecognizable by the immune system will have been at a powerful selective advantage, whereas those B cells that could track the antigenic changes will help protect the host.

In summary, the B cell component of the immune system illustrates two levels of somatic selection. First, the steady flow of new B cells that can react against self components are deleted. Then, when nonself antigen happens to enter the host, those B cells that can react with the pathogen are selected to undergo many rounds of cell division and mutation with repeated selection for those B cells that continue to react with an ever decreasing concentration of antigen. This process is termed affinity maturation. When viewed in the context of the presence of environmental selection pressures (either self or nonself antigens), the individual B cells of the organism undergo a process that is indistinguishable from what is normally thought of as classical gene-based biological evolution of organisms, even though these B cells are not able to behave in all the ways often expected of an organism.

In terms of the overall effectiveness of the immune response, the affinity maturation process is of marginal significance because it occurs after the pathogen has been eliminated and can therefore only act during subsequent reinfections. The two significant selection processes occur first at the level of sorting the stream of new B cells into specificities that are either self (to be eliminated) and nonself (to be kept) and second at the level of amplifying only those B cells with specificities that react with the pathogen that suddenly and unexpectedly appears. The strict notion of serial rounds of replication, variation, and interaction applies only to the small component of affinity maturation in the overall immune response. However, it would be difficult to argue that the immune system does not undergo somatic evolution as a parallel to classical gene-based evolution found in the pathogens.
 

4.3. Somatic Selection vs. Germline Selection

Another important aspect of somatic B cell evolution is whether or not its origin as a part of the developmental program of the host is sufficient to disqualify this process as an example of selection. Included in this question is the inability of the immune system to continue evolving when the host dies. When the host dies of starvation or from being eaten by a tiger, it does not mean that the host's immune system is defective; it just happens to stop evolving because of some unselectable cosmic catastrophe. It seems unnecessarily restrictive to say that selection has to continue for some arbitrary period of time. To be able to show that somatic selection in the immune system stops for some reason other than a failure of the immune system is sufficient to conclude that a process of selection has been at work.

Many similarities exist between somatic evolution in the immune system and the functioning of the nervous system. In contrast to the detailed knowledge of the molecular and genetic structures and functions of the immune system, much less is known about the nervous system, and as our analysis of operant learning will clearly illustrate, even in a well defined behavioral domain, the relevant molecular and genetic factors are almost unknown. Nonetheless, Edelman's ideas on immunology and neurobiology are sufficiently interesting to warrant comment. As a leading figure in the early years of modern immunology, Edelman was a strong proponent of what has come to be termed the "big-bang" version of the generation of antibody diversity. In particular he postulated a somatic genetic recombination mechanism that could generate a huge number of variants without having to resort to point mutations, which he thought to be rare and to occur throughout the genome (Gally & Edelman 1972). This initial burst of genetic diversification dispersed the variants in different B cells which were then subject to selection with respect to self and nonself reactivity. Further genetic diversification after infection and antigen selection was thought to be minimal because it might include the introduction of new specificities able to react with and destroy self-the host.

Under big-bang models all of the diversity of the immune system is generated early in ontogeny, driven largely by the need to eliminate anti-self at an early stage in order to leave the remaining anti-nonself repertoire large and readily induced. This conceptual framework of big-bang diversification in the immune system can be found in Edelman's later work on neuronal development (Edelman 1987). Two points need to be made regarding big-bang and the immune system. First, big-bang diversification necessarily includes all the waste in all of the possible lineages selectable by antigen. Second, Darwinism, if it exists, must surely be more than selection from an unimaginably huge pile of possibilities. As we argue here, Darwinism involves multiple cycles of selection; i.e., interaction, mutation, and replication. Big bang requires only one round of mutation and selection, followed by continuous selection. While big-bang is difficult to justify when the immune system is constantly being regenerated, this is less of a difficulty in the case of the brain, where little cell division occurs once it reaches its adult size.

The purpose for raising the big-bang principle is to emphasize that it denies the kind of serial selection we propose here for the three systems under investigation. Empirical observations notwithstanding, an a priori case can be made that if all possible variants are generated during big-bang and if the fraction of all possible variants used during the lifetime of an individual is very small, then the waste generated by unused variants is prohibitively large. By generating variants among only those cells (or neural connections or organisms) that are already responsive to the selection pressure, waste in the production of unresponsive variants is limited and is not spread among all possible cells in all possible lineages. However, when variation is restricted to those entities responding to a selection pressure, each intermediate variant in a lineage must be individually selectable. If not, then the lineage would become extinct while waiting for a second or third variant to occur.

One underappreciated selection pressure is the relative levels of waste, especially when evaluating probable vs. improbable lineages. Before accepting models of neural behavior based on big-bang diversification followed by somatic selection, it would seem prudent to consider possible alternative models based on serial selection, because the level of waste in the latter is likely to be substantially less than in the former case.

The widely quoted work of Hinton and Nowlan (1987) provides another illustration of somatic selection gone awry. Their assumption of 20 switches in a neural network, each individually inactive, but providing a strong selective advantage when correctly coupled, is close to impossible. It may be true that given this impossible starting condition a form of somatic selection might be envisaged that is capable of selecting the right combination of switches, and that eventually a germline selection for the switches all being in the right configuration is favored because the right combination is always found quickly. However the exercise is rendered moot because the initial assumption is, at best, implausible. There is simply no remote likelihood of 20 gene duplication and mutation steps occurring in the absence of selection of the intermediates (i.e., the intermediate switches from 1 to 19 are individually unselectable).
 

4.4. Population-Level Selection in the Immune System

Several mechanisms have evolved to produce the massive amount of variation necessary to make the immune system work. The genes that code for antibodies have developed a variety of mechanisms needed to rapidly diversify a relatively small number of germline genes in a large somatic population of B cells. Central to these mechanisms is the generation of a functionally haploid genome in the region encoding antibody specificity. The introduction of mutational variants in these haploid specificity regions creates a population of different B cells, which are then subject to further individual selection by antigen. The immune system also exhibits a very different kind of variation that is uniquely expressed at the level of populations of host organisms. This variation is confined to the 2-4 genes that determine what is termed the Major Histocompatibility Complex (MHC)-the locus primarily responsible for the extreme difficulty in transplanting tissues. The MHC genes exist as a large number of alleles (~100) that are found at roughly equal frequency in the interbreeding population. Although there are roughly the same number of alleles of the hemoglobin genes, all but a few alleles are at a such a low frequency that they can be accounted for by mutation alone. To explain the roughly equal frequency of so many alleles at the MHC locus requires postulating a selection process that operates at the level of the genes in individuals and at the level of gene expression in the population.

The exact nature of the selection pressure operating on the MHC genes is not well known, but one compelling, illustrative explanation depends on the role that these genes play in immune protection against viral infections. In order for the immune system to respond appropriately to events occurring inside a virally infected cell (and without cracking the cell open to peek inside), the immune system uses the MHC genes to provide means for transporting intracellular peptides to the surface of the cell. Once these peptides are displayed on the cell surface, a special type of antigen-specific cell (the T cell) is able to bind specifically to the peptide under the right conditions and decide what to do. It works as follows: each MHC allele picks up a slightly different peptide fragment and presents it to the immune system. Viruses may well be selected for if they have mutations that disable or block the peptide binding site on the MHC and so stop the immune system from detecting presence of intracellular virus. To combat this occurrence, the host has at least two, sometimes four MHC genes, each with a different peptide binding specificity. If all individuals in the population had the same alleles with the same peptide binding specificity, then, as the virus moved from one individual to the next, it could keep on evolving to defeat the MHC system. However, if the population possesses a large number of different alleles, then when the virus moves from one host to the next, all the selection in the previous host is canceled because the new host has new peptide binding rules determined by the new MHC alleles. Thus, each allele functions perfectly well in an individual, but selection on the virus extends over many individuals at the population level. One result of this process is the large number of alleles in the population. This situation can be contrasted with the large number of antibody specificities needed per individual. The polymorphism of the MHC locus provides a particularly clear example of selection on alleles of genes that must occur at the population level while still being executed at the level of the individual organism.
 

4.5. Serial Somatic Selection: The Immune System is One Example

In this brief overview of the immune system, we have extracted four examples of selection. (1) In the case of the 100 germline encoded V-segments at the L and H chain loci, these segments can produce 10,000 different LH pairs with different antibody binding sites. The selection that maintains these segments as different V segments is 100 pathogens that would be a threat if it were not for the specificities of this 100 unique LH pairs; the 9,900 other combinations are unselected as particular specificities and represent a very small form of "big-bang." (2) Among the unselected LH combinations and point mutants of these segments (up to one million of them), some are able to recognize self components of the host and have the potential to kill the host instead of the pathogen; these specificities are selected against by killing the cell that makes that specificity of antibody before the antibody is secreted and can kill the host. (3) Some specificities are able to recognize antigens of the pathogens, and the B cells that make these specificities are induced to proliferate and mutate so as to produce new specificities that function better (at lower concentrations of antigen) than others. This example of somatic selection is also an example of serial selection of the type that forms lineages akin to those found in the serial selection processes of the evolution of organisms. (4) In another domain of immune system function, MHC molecules play a critical role in allowing the immune system to be informed of the presence of pathogens located inside the cells of the host.

Although each individual organism (host) has 2-4 different MHC genes, in the population of organisms, there are 50-100 alleles, all at roughly the same frequency; and this implies that selection, which must occur in individuals and their genomes, is via a selection pressure that only affects individuals because they are in a particular population (i.e., the selection pressure is particular to the population an individual finds itself a member of). This process is strictly germline selection, not somatic selection, and is sometimes referred to as "group selection." If we take a hard position on selection processes and require repeated rounds of replication, variation and interaction, then the immune system offers one example of selection in affinity maturation. However, taken together, the other two examples of somatic selection also seem to simulate all of the features we might expect of serial germline selection in classical gene-based organisms.
 

5. Operant Selection
 

In one sense, all the behavior of organisms is the result of natural selection; in another sense, none of the behavior of organisms can be attributed to natural selection. The first statement follows from the fact that natural selection accounts for the range of behavioral potentialities characteristic of the organisms in any particular lineage and also for the processes that account for behavioral content that is uniquely suited to circumstances arising during an organism's lifetime. The second follows from the fact that processes other than natural selection are always involved when behavioral content actually appears in the behavior stream of a living organism. Between these two extremes lies the vast domain where behavioral scientists toil. Although no serious student of science would likely subscribe categorically to either of the two extremes, behavioral scientists with differing interests focus their attention on different segments of the continuum and tend to characterize those with interests elsewhere on the continuum as occupying one or the other of the extremes. Full scientific understanding of behavioral phenomena will require understanding the full range of behavior from one end of the continuum to the other.
 

5.1. Operant Behavior

When the behavior in which scientists are interested changes in content, often dramatically, during an organism's lifetime, one might say that those scientists are interested in the behavior of behavior. While such a locution sounds odd, a cursory look at how "behavior" is used in science reveals that scientists discuss the behavior of volcanoes, proteins, hurricanes, immune system, etc. When change in the phenomena of interest is the object of scientific study, the scientists are said to be studying the behavior of the phenomena. If the phenomena of interest under investigation are the activities of organisms and those phenomena are themselves exemplified by change, then behavior change or the behavior of behavior is the object of scientific study.

The behavior changes of interest here are changes in behavior that occur during a single lifetime. The topic under discussion is further narrowed to those changes in behavior that result from a selection process that is conceptually parallel to the natural selection of organismic characteristics across generations of organisms. Most of the scientists studying this type of behavior designate it as operant behavior, and they designate changes in operant behavior of a particular organism "operant learning." Traditionally, operant behavior has been defined as behavior that operates on the environment and changes over time (in form, organization, or relations to the antecedent environment) as a function of "its consequences." From the present perspective "its consequences" is a shorthand way of saying the "goodness of fit between the behavior and consequent changes in the environment." In short, the particular operant behaviors that emerge and change during the lifetime of individual organisms are the results of "a second kind of selection"-a process that itself is the historical result of the "first kind of selection" (natural selection) (Skinner 1953, 1981)

Many questions arise from a selectionist characterization of operant learning. How does operant learning fit into what we know about the evolution of species by natural selection? How does this "second kind of selection" differ from selection processes that result in the origin (and history) of the species? What are the "units of selection" in operant selection? In this paper, we address these issues briefly. We readily acknowledge that a complete explanation of operant behavior will involve processes other than operant selection, just as organic evolution involves processes other than natural selection. We also acknowledge that not all behavior is operant behavior and, hence, that no discussion of operant learning will answer, or even address, all questions and issues pertaining to the range of phenomena in the domain of behavior.

We have chosen to focus here on operant behavior for both conceptual and practical reasons. Operant processes are known to occur in several phyla, suggesting that their origin reaches deeply into the history of life on earth. Second, and paradoxically, operant selection seems particularly relevant to humans (Schwartz 1974, p. 196). Hominid anatomical features such as opposable thumbs, highly developed cortex, and vocal apparatus may have co-evolved with increasing susceptibility to operant selection. On the practical side, thousands of experiments have yielded a large and complex literature from which to draw, some even conceptualizing results in selectionist terms (e.g. Staddon & Simmelhag 1971). Unfortunately, we can only draw upon an extremely limited part of that literature and will selectively attend to work that clarifies the theoretical perspective presented here. Finally, operant behavior is the area of interest of one of the authors. In the same way that "a zoologist may specialize on vertebrates without denying the existence of invertebrates" (Dawkins 1983, p. 405), we do not deny the existence or importance of behavior that is not operant.

Operant behavior, like biological evolution, is one of those simple topics that are widely, persistently, and sometimes perversely misconstrued or misrepresented (Todd & Morris 1992), thus guaranteeing that at least some readers will find what follows to be at odds with conceptions colored by such misrepresentation. In an attempt to preclude excessive cognitive dissonance, we begin with a few general points that we think critical to understanding the theoretical perspective presented below.
 

5.2. Adaptation and Complexity

In the larger context of biological evolution, an organism's operant behavior has the biological function of interfacing between the organism and its world. An analysis of operant selection requires allowing the organism that behaves to recede to the conceptual background and making the interface itself the object of investigation. This change of perspective amounts to a figure-ground reversal from that which is apparent in direct perception. The new "figure" is by its nature difficult to "see" due to the temporal character of its structure. But behavior has structure of its own. It is made up of parts and wholes, which are parts of more inclusive wholes, and those parts have functions, as do the wholes. An operant repertoire is made up of interrelated behavioral lineages, each having its origin at a different time in the history of the organism, and each having its own history. As in the case of the evolution of life on earth, understanding the process requires focusing on particular lineages. Each behavioral lineage evolves in relation to its local environment, and changes in one lineage can impact other lineages in the organism's repertoire. A particular operant repertoire generally becomes, over time, increasingly complex in terms of the number of lineages it comprises, the complexity of its component interactors, and the historical and ecological relations among them.

The processes by which operant adaptation occurs are viewed here as analogous to the processes by which biological evolution occurs. Specifically, operant selection (in concert with other processes) adapts organismic activity over time to "fit" the environment in which it occurs. If the environment moves out from under the behavior slowly enough, the behavior may be able to adapt to the changing environment. If the environment changes too rapidly, the behavior may be extinguished. As in the evolution of species, operant behavior fits the present environment because of past selection and not because of any future state of affairs. Further, operant behavior that is well adapted to its environment may not contribute to the survival of the organism that is behaving. For example, behavior that is well adapted for producing drug-induced euphoria may result in premature death of the organism. Operant processes work the same way whether or not particular behaviors are conducive to survival of the behaving organism. So far as survival and reproduction are concerned, operant behavior is a very sharp two-edged sword.

Gene-based selection is studied in bacteria and fruit flies as exemplars of a process assumed to account for all species. Similarly, operant selection often is studied in lever presses and key pecks as exemplars of a process that has been shown to operate with respect to more complex behavioral units. Most readers of this paper will readily accept the proposition that a single set of processes accounts for the structural and functional complexity of primates as well as bacteria. Although we trust they can entertain the analogous possibility that a single set of processes can account for structure and function of behavior far more complex than lever presses and key pecks, they may draw their dividing line between operant behavior and "higher" behavior wherever they please.

In the sections below, we provide examples of the ways in which operant selection results in behavior change. The theoretical language used to describe the process is the language we suggest for a general analysis of selection rather than the language used by the original researchers. We readily admit that we are viewing operant behavior from a non-traditional perspective and regret that some parts of the analysis are somewhat speculative. Evolutionary biologists had to develop evolutionary theory for decades in the absence of an adequate theory of heredity. Even after development of Mendelian genetics was under way, considerable time elapsed before these two groups of scientists were able to see how the theories could be combined into a single coherent theory. Operant researchers and neurophysiologists are in a comparable position today. Neural mechanisms are not well understood by most operant researchers and the ways in which operant behavior changes as it undergoes environmental selection are not well understood by most neuroscientists. Although experimental evidence supports a selectionist interpretation of operant behavior and its neural underpinnings, theoretical revision is likely to be required. On the positive side, massive experimental evidence supports a selectionist interpretation of operant learning, and attempts to relate the findings to one another are increasing in number.
 

5. 3. Operant Interactors and the Behavioral Environment

The relation between responses and consequent stimulation (environment) is the area where most operant researchers have focused attention. In operant selection, the primary role of entities traditionally identified as responses is that of "interactor," the unit which "interacts as a cohesive whole with its environment in such a way that this interaction causes replication to be differential" (Hull 1989, p. 96). Although the most obvious entities functioning as interactors in behavioral selection are responses, some interactors in operant selection cannot easily be conceptualized as responses. For example, a group of responses may function as a cohesive whole in operant selection. The members of the group may be homogeneous, such as a burst of lever presses that interacts as a cohesive whole with its environment; or an interactor may be a cohesive whole made up of many different and functionally related parts, as in baking a cake or driving to work. Although the interactors that experimentalists work with in operant laboratories are often lever presses and key pecks, applied behavior analysts have demonstrated in hundreds of studies that operant selection occurs at many levels of behavioral complexity. To assist the reader in relating ideas presented here to previously acquired concepts, we will use "responses" for interactors that can easily be conceptualized as responses and use the more technically correct term "interactors" when the events are less easily conceptualized as responses in the traditional sense.

In operant theory, activity designated as a "response" does not require a stimulus (Skinner 1953, p. 64). Beginning with the assumption that a particular response occurs because it is elicited by another particular event is neither necessary nor helpful. Most operant responses are functionally related to stimulating events, but those relations are exceedingly complex. For purposes of exposition, we will concentrate on the least complex of operant lineages-those that might be compared to prokaryote lineages in biological evolution (Glenn & Madden 1995). Traditionally, these operant lineages have been called response classes, but that terminology raises the same conceptual difficulties that arose from calling a species a class of organisms (Glenn, Ellis & Greenspoon 1992). Both "response lineages" and "response classes," however, imply that operant responses are parts of a population and the characteristics of a population of interactors are the focus of our interest.

When operant behavior is seen as the figure, against organism as ground, the elements involved in selection processes are analogous to (not the same as) those involved in gene-based biological evolution. In operant selection, the interaction step involves a relation between responses (interactors) in an operant lineage and changes in stimulation (consequences) that follow those responses. In the simplest example of operant selection, some relations between behavior and consequent stimulation have the effect of increasing the frequency of responses in the lineage to which the response belonged. This effect is called reinforcement. Other relations result in a decrease in the frequency of responses in that lineage. Depending on the nature of the change in stimulation, this effect is called either extinction or punishment.

The selecting environment (consequent stimulus changes) is a subset of a larger domain of events in the physical world that have function with respect to interactors (responses) in a particular operant lineage. The full range of environmental events having function with respect to the behavior of a particular organism (including events having discriminative, conditional, or motivating functions) is that organism's behavioral environment. Any behavioral environment is a subset of a still larger domain that comprises "the environment" as often construed-the physical world (including that part of it deemed "social"). These different uses of the word "environment" often go unrecognized and are the source of much confusion in the behavioral sciences, as they have been in the biological sciences (see Brandon 1990, chapter 2).

The facts underlying the points in the previous paragraph are incontrovertible, but the conceptual language calls for further explication. The relation viewed here as the interaction step in operant selection must itself be related to the concepts of variation, replication, and retention in operant behavior as one exemplar of our general analysis of selection. In the following sections, those concepts will be discussed in the context of further discussion of the ways in which an organism's operant behavior changes over time.
 

5. 4. Response Frequency in Operant Lineages

One of the earliest and most productive tools of operant researchers was the cumulative record. Although a record depicts only a small amount of information about each response recorded, it captures a critical feature of evolutionary processes-the frequency at which responses in a lineage appear over time. The responses depicted in the record are those that satisfy the contingencies of selection designed by the experimenter. By changing the selection requirements, researchers bring about changes in the frequency, distribution in time, and selectable properties of responses in an operant lineage. Such changes were the initial subject of research on schedules of reinforcement (Ferster & Skinner 1957). Each schedule specifies a particular kind of selection contingency and consequently each results in its own characteristic response distribution.

Although the selection process works at the level of single organisms and results in historical changes in operant lineages in that organism's behavioral repertoire, a schedule of reinforcement produces its characteristic distribution in different operant lineages of particular organisms, across organisms of a single species, and across species on planet Earth. The striking similarities in distributions of responses on a particular schedule in different operant lineages may be viewed as behavioral heteroplasties. That is, selecting environments having particular features result in operant lineages having characteristic distributions. The distributions arise again and again when the selecting contingencies are repeated.
 

5. 5. Selectable Properties of Operant Responses

All operant responses have in common certain properties just as all organisms have in common certain properties. Common properties of organisms include length, width, height, and body mass. Gilbert (1958) identified the "fundamental dimensions" of operant behavior but, from the present perspective, he did not distinguish unequivocally between properties of operant responses and properties of operant lineages. Responses are components of individual lineages in operant selection, just as organisms are components of individual species (and lineages) in natural selection. Some demonstrated selectable properties of operant responses are duration, latency (interresponse time), force, form, direction, and relation to antecedent events. Because the level at which change is measured in selection processes is the population level, the properties of any one response are of little theoretical interest. Evolution occurs at the level of lineages and is measured in terms of response rate (frequency) or frequency of trait values in particular populations of operant responses.

Blough (1963) provided a graphic picture of change in an operant lineage as a function of change in selection contingency. In Figure 1, the response property of interresponse times (IRT) of successive responses in a pigeon's keypecking operant are represented by the height of a dot on the ordinate of the graph. Time is represented on the abscissa. In the first 20 minutes, while a peck produced food only at the end of 4-minute intervals (VI 4 min schedule of reinforcement), the distribution of IRT values in the population of responses was stable. The response population characterizing the lineage at that time shows a good deal of variation in IRT values (0.1s - 6.0s) , with a clustering of IRT values around 0.4s The vertical dashed line shows where the schedule changed to FR 30 (every 30^th peck followed by food). During the next 100+ minutes, IRT values underwent a transition in which more and more IRT values clustered around 0.4s, although variants continued to appear through the whole range of IRT values.

Such a change in population values in an operant lineage is conceptually equivalent to the often cited change in colororation of successive generations of English moths undergoing anagenesis after industrialization. There are differences, however. First, the moth population (as in all sexually reproducing organisms) was distributed in space at any particular time and it extended in time across generations of moths. The distribution of IRT values in successive populations of pecks occurs only in the time dimension because organisms cannot press a lever more than once at a given time. Second, the trait of interest (IRT) in the operant lineage appears to have a broad range of values whereas the trait of interest (color) in the biological population appears to have a small number of discrete values. So at the resolution of human observation, responses appear to vary continuously at least in some of their dimensions.

The formal and temporal properties of responses in an operant lineage may vary widely within a population or they may vary within a narrow range of values. IRTs varied by a factor of 60 in Blough's data (Figure 1). This suggests that IRT may not have been the target of selection. That is, IRTs were not the property ("response trait") on which food was contingent but rather they changed along with other properties that were the target of selection. This distinction pertains to that between selection of multi-dimensional interactors and selection for their particular properties (traits) (cf. Sober, 1984; Glenn & Madden 1995).

Although operant researchers have not traditionally presented data to demonstrate selection for particular properties, Catania (1973) depicted how response populations at successive times could be depicted to demonstrate the effects of selection for response properties of specified values. An adaptation of his graph for present purposes is shown in Figure 2. The changing frequency of force values in a lineage of lever presses is depicted as it could be observed in three populations of responses measured at successive time segments. Force values are fairly evenly distributed in the initial population of responses (A), representing a hypothetical population during a period of time before selection for specified force values. The curve labeled B represents a distribution of force values in a later population, after implementation of a selection contingency in which food pellets follow lever presses having force values between x and y (and not otherwise). The dashed line represents the probability of consequent food for presses having force values between x and y. The distribution of force values in the B population shows the effects of the contingency of selection on the operant lineage. Population C shows the effect of further selection for forces between x and y. In the B and C populations, those responses falling outside the x-y force range represent unreinforced responses (and would not appear in a cumulative record). Although successive populations in this graphical depiction of operant selection are increasingly composed of responses adapted to the selecting environment, variants that fail to meet the selection contingency continue to appear in the lineage, at varying frequencies at different times during the history of the lineage.

5.6. Variation in Operant Selection

Each interactor in operant selection has many properties and each property occurs at various values in responses forming a lineage. Rarely will two or more responses in a lineage be alike in all respects. Interactors with fewer components are likely to resemble each other more closely than will interactors having many parts, each of which can vary along many dimensions. When a behavioral interactor interfaces with the consequent environment, replication of all of its properties either increases or decreases in probability. However, only some of the interactor's properties may be required for an adequate "fit" with the selecting environment (e.g. the force values in the above example). Those properties are differentially perpetuated in the population maintained by the current selection contingency.

Interactors in an operant lineage can be selected on the basis of the property of varying from their predecessor(s). Page and Neuringer (1985) performed a series of experiments in which a sequence of 8 pecks, distributed across 2 keys, was required to differ in their pattern from (a) the previous sequence (Lag 1),(b) the 5 previous sequences (Lag 5), and (c) other previous sequences up to 50 (Lag 50). In the present context, each sequence of 8 pecks is conceptualized as an interactor that either did or did not meet the requirements of the selecting environment. Selection was for interactors with a sequence of parts that differed from the sequence of parts of the interactor's immediate predecessor and for its last 5, 10, 15, 25, and then 50 predecessors. The 8-peck sequences showed variability in sequencing consistent with the selection contingencies, whether the requirements were gradually increased or whether a Lag 50 was implemented immediately (with other, naive, experimental subjects.) Various control procedures demonstrated that interactor variation was itself being selected. In further experiments, the authors provided strong evidence that the variability observed was most likely randomly generated ("the pigeons behaved as a quasi-random generator", Page & Neuringer, 1985, p. 447) rather than the result of some kind of memory function.

There is a difference in variation as a dimension of behavior and measures such as duration or force as dimensions of behavior. Variation is a direct measure of a population (like frequency or rate at which interactors are generated) whereas duration and force are measures of individual members of a population, which can be represented statistically as measures of a population undergoing selection (as depicted in Figure 2 above). Page & Neuringer (1985) concluded that variability in responses in an operant lineage is "a dimension of behavior much like other operant dimensions" (p. 450) in its susceptibility to selection. These and later experiments support that conclusion. And as in other kinds of selection, the susceptibility of variation to selection does not imply that selection is the source of the variation. Variants must occur before selection can operate. Page & Neuringer (1985) suggested that variation is an intrinsic property of operant behavior (i.e. has its origin in natural selection); operant selection can dial it up or down.
 

5.7. Origin of Operant Lineages.

In one sense, as suggested earlier, all behavior has its origin in natural selection or, more proximally, in the inherited behavior of individual organisms. In some cases what is inherited has enough organization to be considered a behavioral lineage. For example, the pecking of pigeons is highly organized, in its formal properties as well as its relation to some properties of the environment, before operant selection begins to adapt features of an individual bird's pecking to local contingencies. Pigeons' pecking is a behavioral lineage that transcends the lifetime of individual pigeons. Its origin is in the history of the species. Operant modifications of the lineage during the lifetime of individual pigeons occur, but they are not encoded in the germline.

Some inherited behavior is not well organized with respect to its environment. Organisms of many species inherit a "supply of uncommitted behavior" (Skinner 1984). It is the kind of activity seen when an organism is in an environment that contains few elements with which it has means of interacting. Such behavior is prominent in the repertoires of human infants and can be seen on occasion in human adults (e.g., profoundly retarded adults or adults submerged in water or isolated for a long time in an empty room). The supply of uncommitted behavior is primordial in phenotypic behavioral development. Operant lineages emerge from the primordial behavior of a particular organism when selection is contingent on particular properties of the primordial activity and, as a result, those properties begin to appear more frequently in the behavior stream. If the selection contingencies gradually tighten, a response lineage gradually forms out of the more or less random (or, at least, poorly organized) activity.

Although the emergence of organized activity from undifferentiated movements can be seen to occur in real time, it has been difficult to study it experimentally because of a lack of equipment that allows recording of both the behavior meeting the changing contingencies and the rest of the behavior in the subject's ongoing behavior stream. Pear & Legris (1987) were able to develop a computer program that continuously tracked the position of a pigeon's head. They specified an arbitrary response (not seen previously during extensive observation of the pigeon in the experimental setting) as the experimental target. The pigeon's head was to make contact with a 3-cm diameter "virtual sphere" at a particular location in the chamber. In addition to its precise spatial location, the form of response to be generated involved a dipping of the pigeon's head at that location. Beginning with a target virtual sphere that the pigeon's head would easily "contact", the experimenters gradually increased the frequency of movements making contact with the sphere and then gradually reduced the size of the sphere. As a result of these changing selection contingencies, the movements acquired the target form and occurred in that form at the target location at high frequency. The interactor lineage that emerged in each of the three pigeons' repertoires was maintained by stable reinforcement contingencies thereafter.

Operant lineages that exist in behavioral repertoires do not all arise from primordial behavior. Many operant lineages come into being by the splitting and merging of previously existing lineages. The complex relations currently studied in operant laboratories involve the merging and splitting of operant lineages (see Sidman 1994 for history of one research program). Such behavioral complexity appears to require interactors that include stimulus parameters, interactors called "stimulus control operants" (Ray & Sidman 1970). The appearance of stimulus control operants in an operant repertoire has been likened to the appearance of eukaryotes in biological evolution (Glenn and Madden 1995). They allow for the grouping of responses into interactors having multiple parts and thus the evolution of behavioral complexity during the lifetime of one organism.
 

5.8. Replication and Retention in Operant Selection.

Selection is a two-step process. "A process is a selection process because of the interplay between replication and interaction" (Hull 1981, p. 40-41). In operant selection, one step is the differential interaction between responses and consequent stimulation (environment) that "causes replication to be differential" (Hull 1989, p. 96). The other step is the differential replication of response characteristics in successive generations. Whether operant selection is a process that parallels natural selection and belongs to the class of theories sometimes called "Darwinian" depends on the requirements one makes of the replication process. If the environment must have multiple and differing copies of a replicator concurrently available for selection to occur, operant behavior seems definitionally excluded. However, there appears to be no reason to assume that all replication processes involve concurrently existing events or objects. All that may be required is a process that retains features of interactors (event or object) across generations in a lineage, with a mechanism of variation to introduce novelty. As we said earlier, successive variations must in some sense be "retained" and then "passed on." This leads to questions regarding the site of retention of operant behavior and the mechanism by which "passing on" is accomplished.

So far as the material world is concerned, what is left after an operant interactor is gone is the central nervous system of the organism whose operant behavior is adapting to changing local contingencies. So the first step in operant selection occurs at the behavioral level (at the interface between organism and environment) and entails relations between interactors in a particular operant lineage and a selecting environment. And the second step occurs inside the organism at the neural level. Research on the biochemical mechanisms underlying learning and memory seeks to identify long-lasting changes that must occur in the strength of synapses as learning progresses and the learned behavior is maintained. "The range of possibilities for memory maintenance is large. None of the proposed models have been firmly excluded, and there seems to be no clear candidate" (Lisman & Fallon 1999). Full understanding of operant selection will require understanding of the relation between the two steps in the selection process. Because one step occurs at the neurochemical level and the other at the behavioral level, such understanding will necessarily entail synthesizing findings from research at these two levels. Unfortunately, researchers working on each of the two subprocesses, like geneticists and evolutionists before the modern synthesis, have little knowledge of one another's findings and often view with suspicion the conceptual framework of the other. There are exceptions. Donahoe and Palmer (1994) have begun to fashion a synthesis of biobehavioral processes in which they view neuroscientific findings in the context of a selectionist theory of learned behavior.

If the site of retention is the central nervous system of the learning organism, understanding of the mechanism(s) of retention will require investigation of changes in the properties (structural or functional) of neural activities as a function of differential interaction between responses and environment. In what follows, we draw on research that explicitly relates replication at the cellular level to operant processes. In a series of publications, Stein and his colleagues set out to assess Skinner's hypothesis that what constitutes a "response" at the behavioral level may not be that which is "strengthened" (i.e., replicated) in operant selection. Rather, a response's "elements" or "atoms" (i.e., characteristics or traits) are the "units of behavior" susceptible to operant selection (Skinner 1953, p. 94 ). No such element can be correlated with a unit of replication smaller than a single neuron, so Stein et al. used in vitro preparations in which single neurons were subjected to analog contingencies of operant selection. For example, Stein, Xue, and Belluzzi (1994) made micropressure administrations of dopamine contingent on spontaneous bursting frequencies of single neurons of a hippocampal-slice. They demonstrated that bursting frequencies increased when dopamine (a chemical associated with the reinforcing effect of drugs) was administered contingent on bursting, that the frequencies decreased when dopamine was not administered contingent on bursting and that bursting frequencies remained at or below baseline when they no longer administered dopamine independent of bursting (non-contingently).

In another study Stein & Belluzzi (1988) injected microadminstrations of dopamine immediately after a postsynaptic neuron was activated by a presynaptic neuron, with a resulting increase in the presynaptic neuron's ability to activate the postsynaptic neuron. Other experiments (Self & Stein, 1992) showed that it was not simply the stimulation of cellular activity that explained the effects of the burst-contingent dopamine. As suggested by Donahoe and Palmer (1994, p. 56), the effects of consequent dopamine on the ability of one neuron to activate another "demonstrates that dopamine can modulate the activity produced by glutamate, which is the major excitatory transmitter at synapses in the cerebral cortex, including those in the frontal lobes."

The work of Stein and his colleagues has several implications relevant to our analysis. First, the in vitro preparation demonstrated that the unit of replication is likely to be only a very small part of "complex neuronal circuitry associated with the reinforced response" (Stein, Xue, & Belluzzi, 1994, p. 156.) A second, related implication is that the combination of cell firings can differ from response to response in a succession of responses of a lineage. Similarly, the combination of genes can differ in a succession of organisms of a lineage. Third, the in-vitro preparation removes the operant selection process from any experiential requirements. As Stein, Xue, and Belluzzi (1994, p. 156) put the matter, "presumably, hippocampal slices do not experience 'highs'"). Fourth, the replication required for retention of interactor properties in an operant lineage may not require (but would not preclude) retention of a string of chemicals (as in DNA) across successive generations. Retention of operant properties in a lineage may instead be characterized in terms of the probability of a neuron activating other neurons of a pathway resulting in effector activity ("synaptic efficacy"). In vitro cellular analogs of operant conditioning further suggest that contingent reinforcement modifies several dynamical properties of a multi-functional network associated with motor behavior and that antecedent stimulation is not required for operant learning (Nargeot, Baxter, & Byrne, 1997; Nargeot, Baxter, & Byrne, 1999). Presumably the mechanisms that account for retention of selected firing patterns will be found in the cellular chemistry of the modified network.

Differential interaction of responses and their consequent environments, then, has the effect of altering probabilities of the firing patterns of neurons. Differentially altered probabilities of events that "pass on information" (in this case, information coding for response properties) may be the hallmark of replication in selection processes. Unfortunately, we know little about the coding of information that is "passed on". Others will hopefully fill this knowledge gap. In sum, differential interaction of operant responses and their consequent environments causes differential replication of the properties of interactors in operant lineages. Researchers who consider themselves to be working in two different scientific domains (the "behavioral" domain and the "neuroscience" domain) have studied, respectively, the two steps of operant selection: interaction and replication. If researchers in both domains were to approach their work from a selectionist perspective and seek to synthesize their findings, a unified biobehavioral science of operant behavior would appear possible.
 
 
 

6. Conclusion.

The goal of this target article was to present a general account of selection that is adequate for three putative examples of selection processes--gene-based selection in biological evolution, the reaction of the immune system to antigens and operant learning of individual organisms. After extensive reworking, sometimes generated by disagreements among the three authors of this paper, sometimes the result of two successive sets of referees' reports, we ended up defining selection in terms of repeated cycles of replication, variation and environmental interaction. These three processes must be so structured that environmental interaction causes replication to be differential.

All three systems include variation. However, the respective amount of variation differs from system to system. For example, point mutations are introduced into genomes at very low rates, but these rates must be too high for selection because mechanisms exist that repair them. Enzymes roam up and down strands of DNA, seeking out "abnormalities" and repairing them. However, in gene-based selection in biology most of the variation is introduced by recombination, not point mutations. Rates of variation are extremely high in the immune system. More than one mechanism exists to make sure that the variation needed for selection is present in ample amounts. How variable operant behavior is depends on how finely we analyze behaviors. The natural units of variation are less obvious with respect to behavior than with respect to the other two systems (Enç 1995).

The most fundamental distinction made in this paper is between passing on information via replication and the biasing of this replication because of environmental interaction. As we have argued at some length, selection is not a single process but composed of two processes-replication and environmental interaction. As a result, the issue of the levels at which selection occurs must be subdivided into two questions: at what levels does replication take place and at what levels does environmental interaction take place? These two questions elicit very different answers, depending on which of the three systems discussed in the paper is at issue. In gene-based selection in biological evolution, replication occurs primarily at the level of the genetic material, while environmental interaction takes place at a wide variety of levels, ranging from genes, cells and organisms to kinship groups, demes and possibly entire species. In the development of the immune system, gene-based selection in biological evolution plays the same role as in any other organismic system, but a second sort of selection also occurs. In somatic selection, those cells that specifically recognize a particular pathogen or foreign body respond and undergo extensive mutation and proliferation. Both replication and environmental interaction takes place at the cellular level. In operant learning, the relation between an organism's responses and consequent stimulation causally affects the organism's central nervous system and subsequent behavior. The net effect is that some responses increase in frequency and others decrease.

Several problems arise in explicating the notion of replication. Even though the notion of "information" is fundamental to any account of replication, we do not provide such an analysis in this paper. We anticipate that in the future, this need will be fulfilled. In replication the relevant information incorporated into the structure of replicators is "passed on" to successive generations of replicators. However, the mechanisms responsible for replication differ somewhat. Although the relevant replication in gene-based selection in biological evolution and the reaction of the immune system to antigens takes place at the genetic level, the details of these processes differ. In addition, the distinction between self and nonself that is fundamental in the immune system does not play a corresponding role in the other two selection processes. In gene-based selection in biological evolution and the reaction of the immune system to antigens, genes replicate by splitting and filling in the appropriate nucleotides. Replication takes place in operant learning at the level of neurological processes, the nature of which still remains largely unknown.

Another difference that emerged with respect to these three instances of selection is between linear sequences of replication and their cotemporal proliferation. In operant learning, organisms react to sequences of events that result in cumulative changes--behaviors are reinforced or extinguished. However, in the other two forms of selection, extensive concurrent variations are presented to the environment. Although we think that a multiplicity of cotemporal replicators massively enhances the strength of those selection processes that incorporate such multiplicity, sequences of replicators that do not proliferate in this way also count as instances of selection, at the very least as a limiting case.

Environmental interaction is also necessary for selection. Some entities must interact with their environments so that the replication processes associated with these interactions become differential. Just differential replication alone is not enough for selection, that is if such processes as drift are to be distinguished from selection. As in the case of information, we were confronted by the problem of distinguishing causal processes from other sorts of processes. In the case of information, none of the current analyses of information make the distinction necessary for selection processes. In the case of causation too many different analyses of causation exist, and none of them are totally superior to all others for all purposes. In all cases, however, selection consists of successive alternations of replication and environmental interaction.

The most common critical response to this paper will surely be that various authors prefer different versions of the three theories than those that we have investigated; e.g., replication occurs at levels higher than the genetic material, the mechanism that we have sketched for distinguishing between self and nonself is inadequate, or they simply do not like operant psychology no matter how it is formulated. But these objections are peripheral to the goal of this paper which is to present a general account of selection that is adequate for the three sorts of theories that we have set out. Alternative versions of these three families of theories count against our analysis only if they cannot be characterized in terms of variation, replication and environmental interaction.

If the preceding discussion has shown anything, it has been how counter-intuitive selection processes actually are. The kind of causality involved in selection processes is, as Skinner (1974) noted, very different from our ordinary conceptions of causation. The two most striking features of selection processes are that they are both incredibly wasteful and yet able to produce genuine novelty and increased levels of organization. Given our ordinary notions, we might be led to ask how such wasteful systems can produce both novelty and increased organization. We suspect that selection processes are able to produce genuine novelty and organization only because they are so incredibly wasteful. The efficient production of novelty and order may not sound like an oxymoron, but we suspect that it is.

Thanks are owed to Marion Blute, Todd Grantham, Allen Neuringer, Peter Richerson, Elliot Sober, David Sloan Wilson, Jack Wilson, and several referees. Langman was supported in part by NIH grant RR07716, and by the Programa PRAXIS XXI (16444), Ministerio da Ciência e da Technologia, Portugal.
 

Andersson, D. I., Slechta, E. S. & Roth, J. R. (1998) Evidence that gene amplification underlies adaptive mutability of the bacterial lac operon. Science 282:1133-1135.

Antonovics, J., Ellstrand, N. C. & Brandon, R. N. (1988) Genetic variation and environmental variation: expectations and experiments. In: Plant Evolutionary Biology, ed. L. D. Gottlieb & S. K. Jain. Chapman Hill, pp. 271-303.

Benson, S. (1997) Adaptive mutation: A general or special case? Bioessays 19:9-11.

Blough, D. S. (1963) Interresponse time as a function of continuous variables: A new method and some data. Journal of the Experimental Analysis of Behavior 6:237-246.

Brandon, R.N. (1982) The levels of selection, In: PSA 1980, vol. 2, ed. P.D. Asquith & T. Nickles. Philosophy of Science Association, pp. 315-323.

Brandon, R. N. (1990) Adaptation and environment. Princeton University Press.

Brandon, R.N., Antonovics, J., Burian, R., Carson, S., Cooper, G., Davies, P.S., Horvath, C., Mishler, B.D., Richardson, R.C., Smith, K. & Thrall, P. (1994) Sober on Brandon on screening-off and the levels of selection. Philosophy of Science 61:475-486.

Brookfield, J. F. Y. (1995) Evoloving Darwinism. Nature 376:551-552.

Campbell, D. (1974) Evolutionary epistemology. In: The philosophy of Karl R. Popper, ed. P. A. Schilpp. Open Court Publishers.

Campbell, D. (1979) A tribal model of the social system vehicle carrying scientific knowledge. Knowledge: Creation, diffusion, utilization 1:181-201.

Catania, A. C. (1973). The concept of the operant in the analysis of behavior. Behaviorism 1973:103-116

Crow, F. F. (1999) Unmasking a cheating gene. Science 283:1651-1652.

Cziko, G. (1995) Without miracles: Universal selection theory and the second Darwinianrevolution. The MIT Press.

Darden, L., & Cain, A.J. (1989) Selection type theories, Philosophy of Science 56:106-129.

Davidson, E. H., Peterson, K. J., & Cameron, R. A. (1995) Origin of bilateral body plans: Evolution of developmental regulatory mechanisms. Science 270:1319-1325.

Dawkins, R. (1976) The selfish gene. Oxford University Press.

Dawkins, R. (1983) Universal Darwinism. In: Evolution from molecules to men, ed. D. S. Bendall. Cambridge University Press, pp. 403-25.

Dawkins, R. (1994) Burying the vehicle. Behavioral and Brain Sciences 17:616-617.

Dawkins, R. (1996) Reply to Phillip Johnson. Biology & Philosophy 1: 539-540.

Donahoe, J. W. & Palmer D. C. (1994) Learning and complex behavior. Allyn & Bacon.

Donoghue, M. J. (1990) Sociology, selection, and success: a critique of David Hull's analysis of science and systematics. Biology & Philosophy 5:459-472.

Dretske, F. (1981) Knowledge and the flow of information. MIT Press.

Edelman, G. M. (1987) Neural Darwinism: The theory of neuronal group selection. Basic Books.

Enç, B. (1995) Units of behavior. Philosophy of Science 62:523-542.

Ferster, C. B. & Skinner, B. F. (1957) Schedules of reinforcement. Prentice-Hall.

Gally, J. A. & Edelman, G. M. (1972) The genetic control of immunoglobulin synthesis. Annu. Rev. Genet. 6:1-46.

Gilbert, T. (1958) Fundamental dimensional properties of the operant. Psychological Review 65:272-282.

Glenn, S.S. (1991) Contingencies and metacontingenci