Learning by Imitation: a Hierarchical Approach

Keywords

Imitation, priming, emulation, hierarchical organization, great apes

Abstract

To explain social learning without invoking the cognitively complex concept of imitation, many learning mechanisms have been proposed. Borrowing an idea used routinely in cognitive psychology, we argue that most of these alternatives can be subsumed under a single process, priming, in which input increases the activation of stored internal representations. Imitation itself has generally been seen as a "special faculty". This has diverted much research towards the all-or-none question of whether an animal can imitate, with disappointingly inconclusive results. In the great apes, however, voluntary, learned behaviour is hierarchically organized. This means that imitation can occur at various levels, of which we single out two clearly distinct ones: the "action level", a rather detailed and linear specification of sequential acts, and the "program level", a broader description of subroutine structure and the hierarchical layout of a behavioural "program". Program level imitation is a high-level, constructive mechanism, adapted for the efficient learning of complex skills and thus not very evident in the simple manipulations used to test for imitation in the laboratory. As examples, we describe the food-preparation techniques of wild mountain gorillas and the imitative behaviour of orangutans undergoing "rehabilitation" to the wild. Representing and manipulating relations between objects seems to be one basic building-block in their hierarchical programs. There is evidence that nonhuman great apes suffer from a stricter capacity limit than humans in the hierarchical depth of planning. We re-interpret some chimpanzee behaviour previously described as "emulation" and suggest that all great apes may be able to imitate at the program level. Action level imitation is seldom observed in great ape skill learning, and may have a largely social role, even in humans.

Introduction

In recent years, many behavioural scientists have come to see imitation as an important manifestation of intelligence in nonhuman species. This is a remarkable sea-change in attitude, since less than a generation ago imitation was regarded more as a nuisance. True intelligence, it used to be thought, is indicated by insight. The "cheap trick" of imitating allowed nonhuman species to simulate intellectual capacities they did not have. Even now, this remains the lay view: imitation may be the sincerest form of flattery, but it is not a sign of intelligence. (A tradition of distinguishing certain kinds of imitation as cognitively complex can be traced back to the last century, as we note below, so the sea-change is in some ways more of a renaissance.) Imitation's recent promotion to the status of an intellectual asset in cognitive science has been accompanied by a wealth of evidence that many nonhuman species cannot learn by imitating the actions they see others perform, whereas even newborn humans are now reported to show imitation. A generation ago, behavioural scientists (like laymen do even today) routinely explained the spread of novel habits among nonhuman species as the result of imitation, but any claim of nonhuman imitation now attracts the closest scrutiny. Most reports can indeed be satisfactorily accounted for by simpler mechanisms of learning ("simpler" in the sense of Lloyd Morgan's scale of complexity, and in involving only mechanisms already thought to be necessary to account for other data). All these simpler mechanisms depend at root on the behaviourist notion of associative conditioning.

We do not question the essential rightness of these changes, but we believe they have not yet led to a good understanding of what the process of imitation involves, and what imitation might be useful for in the lives of other species. In discarding the murky bath-waters of loose definitions and weak scientific control, the baby may have been temporarily lost as well.

In this target article, we review briefly the theoretical apparatus that has been found useful in explaining how nonhuman species can seem to learn by imitating when they are not. Definitions of "true" imitation abound already, but none are universally accepted. Given this state of affairs, we have concentrated on the empirical data that researchers accept (or would accept) as conclusive evidence of imitation, irrespective of how they define imitation. We single out one of the currently favoured diagnostic tests of imitation, the bidirectional control procedure, that incorrectly identifies imitation on the basis of behaviour that can be explained more simply in other ways. Using a common explanatory concept in cognitive psychology, priming, we show that a single mechanism can account parsimoniously for most cases of animal social learning with no need for any "special purpose" explanations. Novelty will prove to be a cardinal requirement of imitation.

With this background, we attempt to develop an approach that can describe the richness of imitative behaviour in humans and some of the great apes. Our aim is to go beyond the question "Is it imitation, or not?", to ask instead "What sort of imitation is it, and why is it used?". Examples of behaviour observed in gorillas and orangutans substantiate our view. These examples are not intended to "prove" that the apes can imitate; that vexed question has been discussed in the original publications and is not the central issue here. Instead, we develop a new way of looking at imitation, arguing that rather than being a distinct "special faculty" it is one of a range of cognitive mechanisms for manipulating hierarchical representations of behaviour. This calls into question the widespread current use of "simple" actions to test for imitative ability experimentally. The adaptive role of imitation may be to help acquire complex, novel behaviour; it may be inappropriate in other situations. Our interpretation has implications not only for animal behaviour and developmental psychology, but for evolutionary aspects of anthropology and neuropsychology. If our approach has merit, it should serve to draw theory in all these areas closer to those of cognitive developmental psychology and artificial intelligence.

1. Non-imitative learning resembling imitation

1.0 Sorting wheat from chaff

The idea that there exists a "scale" of imitative faculties which vary in complexity has existed since the times of Romanes (1884, 1889). The standard belief is that the highest levels of perfection of the imitative faculty are achieved in humans, but that rudimentary forms occur in other species. Various terms have been proposed to capture the difference between this highest level imitation and the simpler processes that generate behaviour merely simulating it. These include "reflective imitation" (vs. instinctive imitation, Morgan, 1900); "imitation" (vs. pseudo- or semi-imitative phenomena, Thorndike, 1898); "sensorimotor stage 6 imitation" (vs. stages 1-5 imitation, Piaget, 1945/1962); "true imitation" (vs. local enhancement, Thorpe, 1956), "Stage 4 imitation" (vs. Stages 1-3, Mitchell, 1987); "observational learning" (vs. other social learning processes, Galef, 1988); "impersonation" (vs. emulation, Tomasello, 1990, citing Wood, 1989).

A host of definitions and criteria have been proposed to sort out the "wheat" of evidence for the special faculty of imitation, from the "chaff" of material that can be explained by other, simpler processes; these have been extensively reviewed and discussed (e.g. Galef, 1988; Moore, 1992; Tomasello et al, 1987; Visalberghi and Fragaszy, 1990; Whiten & Ham, 1992). This approach, of sorting out wheat from chaff, derives from scholars working in the tradition of comparative psychology, especially animal learning theory, who originally took on the task of assessing the capacity for imitation in all nonhuman species. All the resulting definitions of imitation, of which Thorndike's classic "learning to do an act from seeing it done" (1898, p. 50) is perhaps as good as any, are therefore "threshold" definitions that establish only minimal criteria for imitation and thereby portray imitation as a single capacity. None of these comparative definitions address the possibility that, even for imitation in the strong sense of the term, there may exist multiple forms, nor do they aid dissection of what is cognitively involved in imitation. This has been less uniformly true of cognitive-developmental approaches to imitation, e.g. Piaget, 1945/1962, Mitchell, 1987. But although scholars working within this latter tradition have applied their models of cognition to nonhuman primates for the last 20 years, they have only recently begun to consider cognitive processes operating beyond sensory-motor levels. It is perhaps no coincidence that our own backgrounds are in cognitive psychology and information processing; we aim to show the relevance of models of higher level cognitive processes to nonhuman imitation.

Our primary objectives in this article are (1) to set up a heuristic distinction between two different kinds of imitation: copying the organizational structure of behaviour versus copying the surface form of behaviour; (2) to argue that at least the first of these, program level imitation, depends on the organism having the ability to build hierarchical structures of actions, an ability with more general consequences.

We must first follow a more traditional line, however, to set aside the phenomena which we do not wish to discuss from the cases of significance where something new about behaviour is acquired by seeing another individual do it; and to do this, we must undertake a discussion of the other mental processes that can generate copies of demonstrated actions. Extensive mulling over the issue of imitation has generated a long list of hypothetical mental processes which can, independently or in combination, generate copies of demonstrated actions. As the list has lengthened, many attempts have been made to categorize the different possibilities into a systematic and meaningful psychological framework and to standardize terminology (e.g. Galef, 1988; Whiten & Ham, 1992; Heyes, 1993). To our minds, none of the existing classification systems offers a package that clarifies all the important psychological distinctions, or leads to an understanding of how, given the processes that have been identified, genuinely imitative behaviour can best be identified. Terminological proliferation suggests several distinct processes are involved in producing "pseudo-imitation", but we argue that a single process will suffice for most of them. We now offer our own attempt to impose order on the plethora of processes proposed, and on this basis identify phenomena which do not qualify as imitation.

1.1 Stimulus enhancement

The great majority of observations that suggest nonhuman imitation are vulnerable to re-interpretation as stimulus enhancement coupled with individual learning (Spence, 1937; "local enhancement", Thorpe, 1956). Stimulus enhancement is the tendency to pay attention to, or aim responses towards, a particular place or objects in the environment after observing a conspecific's actions at that place or in conjunction with those objects. (In the most powerful formulation, this tendency would be specific to cases where the conspecific is obtaining valued rewards by its actions; see Byrne, 1994.) The result of this narrowing of behavioural focus is that the individual's subsequent behaviour becomes concentrated upon these key variables. Naturally, this increases the chance that the animal will learn to gain the reward it has seen its conspecific obtain, often by performing the same actions; whereas an individual on its own would seldom do so. However, the mechanism which generates this apparent copy is the conventional one of individual trial-and-error learning. Observation of the conspecific's pattern of behaviour is not causal to changing the observer's pattern of behaviour. Here we agree with others in the field, that copying that can be explained in terms of stimulus enhancement coupled with individual learning does not qualify as an instance of imitation.

Stimulus enhancement, however, can be subsumed under a class of mechanism very general in cognitive psychology. A hypothetical example will serve to explain this idea (numerous real cases explained as stimulus enhancement are given in Zentall and Galef, 1988, Moore, 1992, Whiten and Ham, 1992, Byrne, 1994). Suppose a monkey observes another monkey eating under a coconut tree. Stimulus enhancement focuses the observing monkey's attention on the large nuts on the ground under the tree and; it begins to experiment with the nuts and discovers how to crack them open, using actions in its own repertoire; and it will consequently learn more quickly and successfully than if it had come upon the coconut tree alone. It may happen to end up using the same technique as the other monkey: but not because the other monkey showed it. In cognitive terms, the social contribution to this learning process can be rather simply described as priming: increasing the activation of stored internal representations that correspond to those particular environmental stimuli that co-occur with the sight of a conspecific gaining a reward. The concept of priming assumes that there exist structures ("records") within memory that represent familiar or identifiable items (Baddeley, 1990, p.172). For any recognizable locations or objects, this must be the case. Identification of these locations or objects, in the context of a conspecific gaining rewards, increases the "activation" or "salience" of the corresponding records. The "primed" records then channel conventional exploratory behaviour and trial-and-error learning towards the now-salient objects, often producing the semblance of imitation.

1.2 Emulation

Whereas stimulus enhancement changes the salience of certain stimuli in the environment, emulation changes the salience of certain goals. In the simplest formulations (Kohler, 1925/76; Tomasello, 1990, citing Wood, 1989; "goal emulation", Whiten and Ham, 1992), the purpose or the goal towards which the demonstrator is striving is made overt as a result of its actions, and so becomes a goal for the observer too. The observer attempts to "reproduce the completed goal ... by whatever means it may devise" (Tomasello, 1990, p.284). Tomasello et al. (1987) liken this process to "a variant of the stimulus enhancement hypothesis" in which the observer learns something about the environment but nothing about the behaviour of another. How the observer itself reaches that goal is a matter of individual learning or prior knowledge, neither of which is directly influenced by the techniques it has observed. Nevertheless, the observer working towards this emulated goal may well happen by chance to use the same techniques as the demonstrator - hence giving the appearance of imitation. Seeing the actions of the other is not important; what matters is that the concrete result of them is identified, and so can be emulated. Again, we agree with others in the field that such emulation does not qualify as imitation.

In cognitive terms, goal emulation too can be described as a matter of priming. Whereas stimulus enhancement primes brain records of stimuli, emulation primes brain records of goals. All that is necessary for this model is that the goals themselves are familiar or identifiable ones. Primed, activated goals are addressed before unprimed ones.

The meaning of "emulation" has recently shifted, however, to include a wider range of phenomena. This shift in usage is illustrated in interpretations made of the findings from three experiments on imitation in chimpanzees and orangutans. In all three, subjects who observed a demonstrator raking in out-of-reach food with a rake tool subsequently used a similar tool themselves to attain the food; yet they failed to copy some details of the model's technique (Call & Tomasello, 1994; Nagell et al., 1993; Tomasello et al, 1987). Tomasello and his colleagues described all these effects as emulation, arguing that the subjects reproduced the observed goal and learned about the "affordances" of the tool, but used idiosyncratic behavioural techniques to attain it. The affordances of a tool are said to encompass its function as a tool, the fact that the goal could be obtained with the tool, or something about the relationship between the rake and the food (Tomasello et al., 1987, 1993). The meaning of "emulation learning" (Call & Tomasello, 1994) has thus expanded to incorporate observational learning about the properties of objects and potential relationships among them. This sort of learning seems to us a cognitively complex phenomenon open to very different interpretations, and one which may require psychological processes distinct from those which can account for simple goal emulation; we return to this issue in section 2.7, below.

1.3 Response facilitation

In the processes of both stimulus enhancement and goal emulation, the influence of the conspecific on an observer's learning is an indirect one. Its actions direct attention to places and objects but the actions themselves are not copied. Indeed, under some circumstances, the other individual need not even be present to produce the effects. Simply finding coconuts beneath a certain tree, cracked open but still containing a little flesh, may increase the salience of the location and features of coconuts (stimulus enhancement), or stimulate the aim of eating coconut meat (goal emulation). To discover whether specific actions have been copied - imitation - many researchers have resorted to an experimental test. An individual is given the sight of a conspecific (the demonstrator) performing an action of a specific type. The subsequent probability of the test animal performing the same action is then compared with its original, baseline probability of doing so. Imitation is operationally defined as a significant elevation in the frequency of an observed action over the normal probability of its occurrence. In an improved version of this method, two groups of test animals are used, each seeing the same problem solved by a conspecific but in different ways. Then, not only can their baseline frequencies of performing the actions be compared, but the groups can be compared with each other in the frequencies of using each technique (see Galef, 1988, 1992; Whiten and Ham, 1992; Heyes, 1993). Imitation is then defined as a significant divergence between the groups in the frequencies of using the two actions, matching the actions observed. In these animal experiments, the test actions used have always been part of the existing repertoires of the subjects, actions whose spontaneous probability of occurrence is not low.

This experimental technique has also been extensively used in developmental psychology to ask whether very young humans can imitate (e.g. Meltzoff & Moore 1977, 1983). Typically, one of a set of several different target gestures is repeatedly performed by an adult in front of an infant whose responses are filmed. The crucial test of imitation is considered to be a selective increase in the frequencies of matching gestures: for example, significantly more infant tongue protrusion after adult tongue protrusion than after adult mouth opening, and vice versa for mouth opening. Positive results have been confirmed in many laboratories, and it is now accepted that several different facial gestures are copied by infants, even when tested only a few hours after birth. The simple gestures have sometimes been "novel" ones for the very youngest infants, in the sense that they have not yet performed them, unlike the case in the animal work. Nevertheless, the spontaneous probability of these actions occurring is not low, so the actions are evidently in the (as yet unexpressed) repertoire of the neonate. The researchers probably have little choice in this, since the neonate or weeks-old babies used in the experiments lack the ability to copy a wide range of gestures.

There is now a mutually supportive consensus among many developmental and comparative psychologists that this sort of experimental paradigm is conclusive evidence of imitation. By this criterion, the ability to imitate has now been detected in a few species of animal (e.g. budgerigars, Dawson and Foss, 1965, and Galef et al, 1986; rats, Heyes and Dawson, 1990, and Heyes et al, 1992; chimpanzees, Whiten et al, 1996), and in near-helpless neonatal human babies. However, we dispute that any of these experiments, using either animal or human subjects, provide evidence of imitation. In the case of animal data, we again propose priming as the explanation.

If the salience of stimuli can be increased, and goals can be foregrounded, by observation of a conspecific's actions, then surely an individual's responses might also be facilitated by what it sees? In cognitive terms, just as brain records of stimuli and goals may be primed or activated by observation of others, so those of responses might also be primed, making them more likely to occur (Byrne, 1994). If an individual saw another gaining a reward while performing a response that physically resembled one in its own repertoire, then the corresponding brain record would be primed and a matching response made more likely in its own subsequent behaviour. As before, all that is required is the existence of structures in memory corresponding to the facilitated actions, which has to be the case for actions in the existing repertoire. Like stimulus enhancement and goal emulation, this simple phenomenon of "response facilitation" would simulate imitation under some circumstances (Byrne, 1994). It is important to distinugish this theoretical proposal from two different phenomena. In "contagion" (Thorpe, 1956), actions that an individual sees performed by another may trigger the same actions in the observer - as in contagious yawning. But here the linkage is innate and involuntary, whereas in response facilitation the effect may potentially occur with any action in the individual's repertoire, voluntary as well as involuntary, provided the individual's perceptual system registers the physical resemblance. In "social facilitation" (e.g. Bandura, 1986; Galef, 1988), the motivational homogeneity among a group of individuals is increased and they may thus tend to perform the same behaviours at the same time, but no specific performance of a motor act is influenced, as happens in response facilitation.

Any experiment that uses changes in the relative frequencies of actions already present in the individual's repertoire as evidence of imitation is potentially vulnerable to re-interpretation as response facilitation (Byrne, 1994; Byrne and Tomasello, 1995). Of course, if an action appears radically different in form when viewed from the perspectives of demonstrator versus performer, or is invisible to its performer as in the case of the tongue-protrusion used in neonatal human work, this criticism has no force. Priming can only apply once the identity is registered, and the very means of recognition is what requires explanation in neonatal imitation. (It remains possible that contagion could account for babies' matching of maternal gestures, since only a few gestures are involved (Anisfeld, 1991); close behavioural matching between mother and infant is potentially beneficial for the relationship, so the evolution of innate linkages for a few discrete actions is not implausible.) In contrast, the animal work has employed responses that look essentially the same to the test animal when performed by itself or another. We believe that this sort of experiment which relies on existing responses is in principle insufficient for any convincing demonstration of imitation in animals or humans (see Byrne and Tomasello, 1995, for detailed critique of one claim to the contrary).

1.4 Implications of priming

We would thus unify three apparently different phenomena and explain them all by a single theoretical mechanism, one already found to have explanatory power in cognitive psychology where it is extensively used. In contrast to the several supposed mechanisms that are sometimes invoked to explain behaviour mimicking imitation (and some behaviour claimed to be imitation), we propose a single mechanism of extreme simplicity, thus reducing the amount of "special purpose" theory needed to understand behaviour. When this mechanism, observational priming, operates on records of stimuli in the immediate environment, responses in the individual's repertoire, and goals which it might choose, the result can look very like imitation indeed (as discussed in Byrne, 1994).

Priming can never produce an entirely new behaviour, or an entirely novel arrangement of "old" behaviours. Novel behaviour can only arise from other processes, possibly enhanced by priming: individual trial and error learning, insightful planning and thought, or imitation. Novel behaviour that arises from trial and error will have a characteristic signature in the history of reinforcement. Novel behaviour might arise from insightful planning and thought, but this we take to be an even less parsimonious explanation than imitation where the latter is also possible, as in all cases discussed in this paper. Therefore, we argue that novelty of behaviour is an essential part of any proper definition of animal imitation. In many past definitions the acquisition of new behaviour was not stressed or was seen as only one aspect of imitation, such as Thorpe's "copying of a novel or otherwise improbable act". We support Zentall's firm ruling that to be sure of imitation, the act should not already be part of the animal's repertoire, whether improbable or not (Zentall, 1996). Humans, as we know from conscious experience, sometimes imitate actions that they can perform already; but if this were shown by animals, other explanations for their behaviour would always remain possible.

2. Imitation by animals

Turning now to manifestations of social learning that appear to be cognitively more complex and not explicable on a simple priming and trial-and-error model, we will first consider how animal behaviour should be properly described. We argue that, at least for certain species of animal, behaviour is hierarchical - at levels that allow functional control by the individual, not merely in the underlying organization of units to which the individual has no access and no possibility of control. For these species, it is a real issue to determine at which level or levels in the hierarchy of behaviour imitation would take place, if it were to occur. The case that some animals can control the hierarchical organization of their behaviour will be made with data from the routine food-preparation activities of mountain gorillas, engaged in their daily activity of eating plants. We aim to establish that great apes have the possibility of imitating behaviour at an organizational level of description, the "program level". Then we go on to use behavioural records of orangutans, engaged in attempts to copy the actions of human caretakers, to dissect program level imitation into its components, most crucially the observational learning of how to manipulate object-object relationships. Finally, we apply these ideas to a better interpretation of some existing data from chimpanzees. Our primary aim is not to attempt to convince sceptics that great apes can imitate. While we do consider this to be much the most parsimonious interpretation of current evidence (and we each made that argument in the original data papers), here our intention is to propose and defend a new interpretation of how imitation works and what it is for. We suggest that imitation, in the sense of acquiring skills by observation, can best be recognized by its organizational structure; that its biological function is to allow observation to be used (in conjunction with other methods) to facilitate the building up of novel complex, hierarchical organizations of simpler units of behaviour; and that only species with control over the hierarchical organization of their behaviour can in principle imitate in this way.

2.1 Hierarchical organization of behaviour

It has long been hypothesized that behaviour is hierarchical in organization. Lashley (1951) argued that the linear serial order of actions concealed an underlying hierarchical structure, and this structure rendered stimulus-response models inadequate; the issue for Lashley was hierarchical structure that is under some voluntary control. More recently Dawkins (1976) proposed hierarchical organization as pivotal to understanding the evolution of behaviour, arguing by analogy with many other cases in developmental and neural biology which had already been found to be hierarchical, and on grounds of efficiency. Hierarchical organizations of control, he showed, are easier than linear ones to repair when they fail, allow the economy of multiple access to common subroutines, and combine efficient local action at low hierarchical levels while maintaining the guidance of an overall structure. In human behaviour, hierarchical structuring has long been argued to be essential for many acquired skills, such as language, problem solving and everyday planning (Chomsky, 1957; Newell, Shaw and Simon, 1958; Miller, Galanter and Pribram, 1960; Newell and Simon, 1972; Byrne, 1977). And theories of cognitive and linguistic development argue that age-related increases in cognitive and linguistic complexity are the products of hierarchical mental construction processes (Case, 1985; Gibson, 1990, 1993; Greenfield, 1991; Langer, 1993).

The proposal that behaviour is hierarchical may apply in at least three ways. First, evolution may tend to favour hierarchical structuring, as Dawkins has argued, in behaviours whose execution is under tight genetic guidance. Second, automatic processes of learning may organize groups of behaviours, learnt singly and linearly, into hierarchical structures. In animal learning theory the possibility of "second-order" or hierarchical association has sometimes been proposed, for example between a discriminative stimulus and a response-reinforcer relationship (Rescorla, 1991); however, even this limited amount of hierarchical organization is not fully accepted (Mackintosh, 1994). In neither of these first two cases does the individual organism have voluntary access to, or control over, the hierarchical structure; the structures are simply induced and triggered by constellations of stimuli.

We wish to consider a third and more radical possibility (already entertained by Lashley), that individuals of some species of animal have access to the hierarchical structure of their own behaviour, and control over its organization - just as humans show in planning or problem-solving. In distinguishing this case from the first two, the crucial issue is whether behaviour is controlled by an elaborate but modifiable structure of goals and sub-goals. If so, then the interesting question for imitation then becomes the extent to which individuals can and do imitate this organisation of behaviour, rather than the old issue of whether they can or cannot imitate a particular action. The answer to this new question impacts on the nature of intelligence in spontaneous behaviour. We would be reluctant to describe as intelligent any sequence of behaviour whose mental organization was a single unit of action connected to a goal-representation, a long sequence of linear associative connections, or a rigid hierarchical structure. Thus, whether a behavioural structure is modifiable by the individual becomes crucial in diagnosing it as "intelligent".

We now turn to two "case studies", on gorillas and orangutans, that we consider point to Lashley's proposal as applying in this most radical form, at least in the great apes.

2.2 Case-study: Hierarchical order in mountain gorilla food preparation

Gorillas inhabit rainforest over a wide area of west and central Africa. As in other great apes, ripe fleshy fruit forms a major portion of their diet; and as with other apes eating fleshy fruit, their techniques of eating fruit are typically not subtle or complex. However, gorillas have also colonized a quite different habitat, and one small population inhabits the subalpine moss forest and temperate meadows of volcanoes in Rwanda, Zaire and Uganda. In these habitats, there is almost no fleshy fruit, and these mountain gorillas feed instead on the leaves and pithy stems of herbaceous plants (Watts, 1984). This diet is nutritionally superior to relying on ripe fruit, as it is rich in protein and trace elements (Waterman et al, 1983); this largely obviates the need for the supplementation with animal and fungal matter which is seen in lowland gorillas, orangutans and chimpanzees. However, the favoured leaves and stems are protected by physical defences like spines or stings, or encased in hard and indigestible material. The only other mammals that regularly eat these plants are ungulates, which have stomachs that are either specialized or very large - in both cases allowing bacterial digestion of plant matter. Gorillas, like other great apes including humans, have simple stomachs. Gorillas cope with the problems in quite a different way, using manual skill (Byrne and Byrne, 1991, 1993).

In describing gorilla feeding, we will concentrate on just two plants, nettle Laportea alatipes, covered in painful stings, and bedstraw Galium ruwenzoriense, covered in tiny hooks that enable the plant to clamber. The gorillas' techniques for preparing nettle and bedstraw to eat appear to be adjusted to minimize the unpleasantness caused by these physical problems. For nettle, their technique effectively removes the worst stings, those found on the main stem and leaf petiole, and enfolds others on the leaf margins. A flow-chart representation of the process (Figure 1a) emphasizes the need for bimanual co-ordination, as well as the overall complexity. After pulling a plant into range, one hand is formed into a cone shape, cupping the base of the stem, and swept upward, stripping a whorl of leaves from the stem. This process may be repeated, while holding the already-stripped whorl(s) with the lower fingers as another whorl is obtained, until a good handful is ready. Then, the leaf blades are firmly gripped, and with the other hand the petioles are grasped; the two hands are twisted or rocked against each other, detaching the petioles which are discarded.

As an alternative to simply repeating the process of stripping, the whole of the first few stages may be iterated, again holding the bundle(s) of prepared leaf blades with the lower fingers while another bundle is added, until a good handful is ready. If there is any dry or otherwise inedible debris, it is then picked out from the mass of leaf blades held in the half-open hand. Next, the handful of leaf blades is partially pulled out from the closed hand, folded over the thumb, and grasped again - forming a "sandwich", with the powerful stings of the leaf margin enclosed within a parcel that presents to the outside only the less sting-infested under-surface of the leaf. This parcel is then popped through the sensitive lips, without contacting them. By this means, all major areas of stings are removed or enfolded, and a substantial handful of nutritious leaves is eaten at once.

The technique for processing bedstraw (Figure 1b) is similarly complicated, but has different functions. Here, the problem is that tiny hooks on trailing stems would tend to catch on throat and mouth surfaces, making eating inefficient and giving a risk of choking. The gorillas' technique here works by compressing a bundle of stems to eat with slicing bites, so that the tiny hooks cannot attach to the interior of the mouth. Once a mass of bedstraw is hauled into range, tender green stems are picked out, the picked stem(s) being repeatedly held with the lower fingers while others are added, until an adequate mass is built up. Loose stems are folded in to the bundle, both during and after this iterative accumulation, either using the other hand or - if the gorilla is in a tree and that hand is already employed in holding on - by rocking the hand back and forth, grasping and re-grasping the stems that gravity causes to fold. Like nettle, bedstraw often includes dead leaves and other debris that must be removed before ingestion; and where this is the case the actions of removing these inedible items are identical in both techniques. The manner of ingestion, however, is quite different: the bundle is rolled against the chin or hard-palate to compress it, then sliced with shearing bites of the molars, rather like chopping herbs on a board.

Since these techniques are found only in a small population of gorillas, and indeed are valueless outside the very limited altitudinal zones in which temperate plants like nettle and bedstraw grow in Africa, there is no serious doubt that they are learned. But is their structure the sort of thing that an associative mental process could produce, or are there signs of complex hierarchical organization under voluntary control? No field experiments have been, or could be, carried out on the members of this highly endangered subspecies of gorilla. Fortunately, however, food preparation is not a rare activity, and the data from hundreds of hours of focal-individual observation are available to help understand the structure of the techniques.

For a start, we know that at the most detailed level of description of manual actions, each individual uses several variants of each behavioural element in the process (Byrne and Byrne, 1993). These variants apparently have no functional significance - each works just as well, and the variant chosen is most likely partly determined by the environment, the plant itself - just the sort of low-level, local decision making that Dawkins pointed out as an advantage of hierarchical structure. An associative, probabilistic process may well be an adequate representation of this process of selection. (However, a choice hierarchy cannot be ruled out; see Dawkins, 1976, for the criteria which enable a choice hierarchy to be distinguished from a Markov model.)

Turning next to the processes' structural organization, does this show any clearer signs of hierarchical structure, of the sort that the individuals can control voluntarily? The strongest evidence that it does is given by the animals' ability to iterate parts of each process, already noted. Miller, Galanter and Pribram (1960) first pointed out the importance of iteration till some pre-defined criterion was reached, a test-operate-test-exit loop. A loop is not itself a hierarchical structure, but it betrays the presence of a subroutine which can be iterated. Figure 2 uses this evidence to show the hierarchy that is implied for processing nettle. The presence of optional processes (those shown in brackets in the flow chart of Figure 1) serves both to parse the string into definite components, and further shows that a linear chain would be an insufficient description, since one link in a chain cannot be "dropped" without losing the place in the sequence altogether. The tree structure of Figure 2 displays the minimum necessary complexity that is implied by the observed presence of loops and optional processes; in reality, there may well be further hierarchical organization accessible to modification in some of the "single act" stages, but we have no evidence for this at present.

In linguistics, re-write rules are often used as a compact way of representing the branching tree structure that logically results from a goal hierarchy, and the nettle technique can easily be represented as a phrase-structure grammar of goal and sub-goals:

nettle find + collect + (clean) + fold

collect strip + tear-off + (collect)

strip strip + (strip)

When a goal hierarchy of this nature is expanded, the system must somehow keep track of where the process has got to; this requires a working memory. As the problem is "unpacked" into component subgoals, the system needs to be able to re-trace up the hierarchy, after dealing with one subgoal, to proceed with subsequent ones; some memory has to keep track of where to return to. In a computer, this is often done with a push-down stack. As search delves deeper into the branching structure, subgoals are added to the top of the stack; when one goal is satisfied (popped off the stack), the next is automatically available (pops up). The size of the heap on the stack varies with depth of unpacking. In a real system, such as the mind of a gorilla or a human, short-term memory capacity might set a limit on how deep a hierarchy it could expand without getting in a muddle.

With a physical task, as opposed to that of constructing a sentence, the "external memory" of the current state of problem may relax this capacity constraint on working memory; and for gorilla feeding, the state of a part-processed handful is continuously in view. This makes a production system, in which productions are selected successively by the current state of the problem, a particularly appropriate representation of this sort of behaviour (see below, and Byrne, 1995). A linear, chain-like process, by definition, has no equivalent mechanism for jumping from one point to another. If a routine were shared between two food-processing sequences - as is the case in gorilla feeding for the delicate operations involved in cleaning out debris from leaf bundles during preparation of nettle or bedstraw - confusion must be expected (see Figure 3). Were such a chain-like model to be correct, errors should sometimes occur, of a distinctive sort: for instance, suddenly treating a bundle of nettle leaf blades as if they were bedstraw stems, and rolling the bundle against the chin. Nothing of the sort occurs. The processes of mountain gorilla food preparation are simply not well described as linear and chain-like: their organization is hierarchical.

"Hierarchical organization" should not be taken to imply one in which control is always imposed "top-down". As noted above, it is likely that many decisions about manual actions are local ones, dependent on the precise detail of plants encountered. All that we mean is that the organization of goals and sub-goals is hierarchical; the sub-process working towards each sub-goal may well have considerable autonomy, so there may be periods of "heterarchical" control. Nor do we suggest that the hierarchical organization of skilled manual actions by gorillas is in all ways similar to human action in skilled manual tasks. For instance, there may be a quantitative difference in depth of hierarchical expansion. Gorilla food preparation appears organized into shallow hierarchies, and perhaps gorillas can only keep track of two embedded goals even with the aid of the "external memory" present in physical tasks.

Finally, we would emphasize that we do not suggest that gorillas' organization of behaviour is inflexible in structure, like some conventional computer programs. The evidence that we have used, to show that gorillas have functional control over hierarchical structures of actions, shows that such rigidity is unlikely. Representation of their techniques as production systems captures this sensitivity nicely: if a stage is unnecessary, the corresponding sub-process is never evoked (see Figure 4).

2.3 Hierarchical levels of imitation

Having illustrated the hierarchical nature of gorilla techniques of plant feeding, we turn to the implications for imitation. How might the complex techniques used by mountain gorillas be acquired? There seems little doubt that learning what to eat is straightforward for infant gorillas, since they are showered with food remains by the mother from the first day of life. In addition, the priming effect of stimulus enhancement would tend to focus a young gorilla's attention on the plant-as-growing as a potential object to investigate. The interesting question becomes, how does a gorilla first acquire the elaborate sequence of co-ordinated actions that converts, say, nettle plants to edible mouthfuls? The answer must be consistent with two facts. After 3 years of age, the age of weaning, there is no further change in efficiency as measured by time to prepare a handful of food (Byrne and Byrne, 1991): techniques are learnt quickly. Secondly, for an infant mountain gorilla the only potential direct social influences on feeding are the mother, who an infant usually accompanies when foraging, and the silverback leader male. Other individuals are intolerant of the presence of nearby conspecifics when feeding, and the dense herb vegetation means that animals are out of sight when only a few metres apart.

Unlike the imitative behaviours usually studied by comparative psychologists, learning a new gorilla feeding technique is not a matter of adding a unitary action to a limited repertoire. Instead, many acts must be built up into one particular logical structure, out of a vast range of potential structures (the term "structure" is used to include sequential regularity, bimanual co-ordination, and the organization of subroutines). The novelty lies in the arrangement, and the skill is to arrange some basic repertoire of actions into novel and complex patterns rather than to learn new basic actions. The lowest level in the hierarchy would consist of the smallest possible elements of action which can be independently controlled by the individual. In a great ape, extensive neural representation of the hands allows rather precise control, and this set of basic elements may be very large. Integrated groups of these basic elements, assembled together during the animal's interactions with the world, form higher-level units of behaviour, and this process of hierarchical grouping can continue to arbitrary complexity, though in practice it may be limited by the available mental capacity. The repertoire of the individual consists therefore of all the lowest-level elements plus the already-assembled higher-level groups of elements, because each of these behavioural complexes functions as a single unit once it has been learnt. Novelty is found in those patterns of behavioural units (including basic elements and already-integrated groups of elements) which are assembled for the first time. In linguistics, it is a commonplace that the multiple levels of patterning in speech - distinctive features, phonemes, morphemes, words, sentences - enable a discrete, and indeed quite small, set of features in speech to encode an almost infinite range of utterances. The mechanism of this "productivity" in speech lies in the hierarchical structuring of groups of elements, in novel orderings and circumstances. Following Lashley, we suggest the same applies to skilled action.

In a hierarchical system, the way in which the process of learning new skills can be aided by imitation becomes similarly a less straightforward matter. With hierarchically structured behaviour, there exists a range of possibilities for how imitation might take place, beyond the simple dichotomy of "imitation" versus "no imitation". Imitation allows the assembly of novel sequences of units by observation; but - given the possibility of several degrees of hierarchical embedding - imitation might occur at many different levels, with radically different consequences for what we would observe. Figure 5 illustrates in words a number of possible levels at which the gorilla skill of processing bedstraw might be copied. At one extreme, the resulting behaviour would be indistinguishable from goal emulation: bedstraw is chosen as a goal, after watching another individual eating it. At the other, comparative psychologists would be confident that imitation, in Tomasello's (1990) sense of "impersonation", was occurring throughout, because the details of manual style, hand preference and idiosyncratic movements would closely match in the behaviour of model and observer. The emphasis on rather exact duplication of the detail of behaviour is an inevitable result of comparative psychologists' use of simple, non-hierarchical actions in their experiments (Custance, Whiten & Bard, 1995; Heyes & Dawson, 1990; Tomasello et al, 1987; Whiten & Custance, 1996). However, it is also possible that the overall form might be imitated, but the fine detail acquired by trial and error. This would result in a striking match of behaviour at coarse levels of description (i.e. higher hierarchical levels), contrasting with non-matching actions at fine levels of detail. Or, some subroutines might be imitated within an overall form that is independently constructed. Thus, imitation operating at different hierarchical levels would produce distinctively different patterns of behavioural similarity and difference at different levels of analysis. What pattern do gorillas show?

2.4 Program level and action level imitation

Beginning at the bottom of the hierarchy, there are two reasons to think that the precise details of the manual actions an individual uses are learnt without imitation. At a fine-grained level of description, where each element of behaviour is distinguished by an exact hand-configuration and movement pattern, each animal was found to have a different preferred set of functionally equivalent variants (Byrne and Byrne 1993). This striking idiosyncrasy is a hallmark of trial and error acquisition. Moving up to a slightly coarser level, at which minor style differences are ignored but left- and right-handed forms remain distinct, the pattern of hand-preferences (which are very strong in every animal) can be used to trace any copying. Since only the mother or the male leader are potential models, if this level of detail were learnt by imitation, hand preferences would strongly tend to run "in families". The hand preference of an offspring would correlate either with that of the mother, or that of the silverback male. In fact, they correlated with neither: their distribution was just what one would expect by chance (Byrne and Byrne 1991). At these low levels, there seems no need to invoke imitation as an explanation.

A very different pattern emerges when we move up to the level of overall form of the process. Once sets of low-level elements, each member of which achieves the same function, are lumped into classes, the variability of technique largely vanishes. At this level, instead of idiosyncrasy, the sequence of actions is a rather fixed one; and in every animal, the organization of each technique is essentially the same (Byrne and Byrne 1993). Even treating right- and left-handed versions of the exact same sequence of acts as "different", animals were 70-80% reliant on just one technique for each species of plant. Given the very large number of possible sequences in which the 6-10 different sub-processes could be combined, this standardization is remarkable, and contrasts strikingly with the variability we found among element repertoires. Of course, many of the theoretically possible sequences do not succeed in processing the food; but plenty do. Even the very simplest technique, structurally unrelated to the adult gorilla's method, (picking off nettle leaves, one by one, and eating each leaf blade while holding the stalk) allows feeding, albeit slowly. Environmental influences can mould regularity from trial and error learning, and some of the more "absurd" sequences of actions could quickly be rejected; but it is highly implausible that the constraints of the environment should be so tight that every animal would end up with the same hierarchical structure, yet so weak that the fine details of the techniques are highly variable between individuals. It would be convenient if there just happened to exist, distributed among the gorilla population, two or more equally efficient techniques for processing a plant; then we could observe if techniques, unlike laterality and details of hand configuration, ran in families. However, this situation is unlikely in any natural ecosystem (in fact, we believe it to be unknown), and the harsh montane environment of these gorillas would exact a heavy toll for feeding inefficiency. As it stands, the sharp difference between individual variation at one level of organization and group consistency at a higher level, in a learned behaviour pattern, leads us to suggest that gorillas may indeed be able to imitate at intermediate hierarchical levels, effectively copying structural organization but not minor details: a kind of observational learning one of us has described as "program level imitation" (Byrne, 1993, 1994). Learning by individual experience is not completely disproven by these data, but it becomes a contrived alternative.

Program level imitation may be defined as copying the structural organization of a complex process (including the sequence of stages, subroutine structure, and bimanual co-ordination), by observation of the behaviour of another individual, while furnishing the exact details of actions by individual learning. We refer to this as the "program" of behaviour as it makes up a recipe for co-ordinating and scheduling acts, and when it is enacted some result is produced by the individual. Imitation at the program level, then, would consist of copying a novel arrangement. By observation of an individual which already possesses a certain program, the observer, using elements already in its repertoire, learns to replicate the sequential regularity and co-ordination of elements and any sub-routines and loops in the flow of control (Byrne & Byrne, 1991, 1993; Byrne, 1994).

The process by which program level imitation might be achieved is not necessarily a mysterious one. In the mountain gorillas' plant processing techniques, the crucial sub-goals are visible (by their results) in the sequence of action: these states necessarily recur in every effective sequence, whereas the irrelevant details of precisely how each of these states is achieved will vary between occasions without affecting efficiency (see Figure 6). Thus an individual observing a skilled model would be in a position to identify not only the final result to aim for, but also the appropriate sub-goals on the way. The detail of how each sub-goal is met can be acquired by individual learning, a process that may in this instance be much more efficient than the imitation of all the movements themselves. To imitate in this way, the individual must have mental apparatus that allows hierarchical frameworks to be assembled, in order to organize this goal structure, and to hold the goal structure while its detailed enactment is built up. Novel frameworks could, of course, also be assembled on the basis of trial and error exploration, and no doubt in simple cases this is quite sufficient; but imitation confers benefit in boosting the rate of acquisition - important where long sequences produce "combinatorial explosion" - or introducing features unlikely to be invented independently.

The results expected from program level imitation are quite different from those sought by comparative psychologists as evidence of "true imitation" or "impersonation", in which a novel action is added - as an unmodified whole - to an individual's motor repertoire (Tomasello, 1990; Whiten & Ham, 1992). The hallmark of this sort of imitation is that style and minor details should match between mimic and model; for consistency, we call this action level imitation. Since (almost) no theorist would wish to restrict "imitation" to the observational copying of single muscle twitches, it might be argued that the only possible (detectable) imitation is program level imitation (e.g. Whiten et al, 1996). This is an empirical question, however. Action level imitation might occur, for instance, by a process of kinaesthetic-visual matching, in conjunction with cognitively simple (associationist or sensori-motor) processes, as envisaged by Mitchell (1993, championing an idea originally proposed by Guillaume, 1926). This could certainly copy a sequence of acts, each of which organizes a complex sequence of muscle twitches, but the contributions of action level imitation would remain entirely linear in structure. Imitation at program level necessarily implies an intelligent ability to operate with task structure and hierarchical organization of behaviour. Lacking this organization, action level imitation might be parodied as the sort of imitation a video-recorder is good at: exact duplication of details and stylistic quirks, but without any implicit understanding of organization.

We suggest that the everyday usage of "imitation", and the sense often used in traditional developmental psychology, are closer to program level imitation. Consider, for instance, the imitation described by Bauer and Mandler (1989) when they studied 16 month old children's copies of adults' sequences of action. They found that the children required less observation and were more accurate if the sequences were of actions that were causally related, such as the steps in bathing a teddy-bear. Any actions unrelated to the job in hand tended to be missed out in the children's imitations. While the individual elements of action are unlikely to have been new for the children, the task's structural organization - not simply an ordered string of actions - was copied, characteristic of program level imitation (see also Abravanel and Gingold, 1985).

In section 3.0, we will consider whether humans make much use of action level imitation, and if so, what for; Wood (1989), who introduced the term "impersonation", noted that some children go on to become stage impersonators or actors in adult life, implying that the faculty of action level imitation has a limited usefulness. Nor is the empirical discrimination between action- and program level imitation necessarily straightforward. When there are clear signs of idiosyncratic, individual learning at lower levels of organization, as in our gorillas, action level imitation can be ruled out; but if this were less clear-cut, then the observable behaviour, whatever its true origin, might look very like the result of action level imitation.

Despite these reservations, many recent experimental and observational studies investigate whether animals imitate by focusing on the action level (impersonation). As we and others have amply charted, unequivocally establishing this fact is very difficult, because of the difficulties of excluding stimulus enhancement, response facilitation and goal emulation as alternative explanations for putative cases of imitation.

The problem becomes more, not less, tractable when species with very flexible action repertoires are considered, such as the great apes. Indeed, the whole concept of an action "repertoire" may be inappropriate for these species, implying as it does a fixed and enunciable list of discrete actions. As humans, we accept that folding, twisting, pulling-apart and squeezing-together are "single actions" - but each is composed of more elementary motor movements programmed together to achieve their results. In turn, the integrated behavioural complexes of pulling-apart-while-twisting, and squeezing-together-while-twisting, once they are acquired (perhaps by trial and error), will function as single units of behaviour in future. It is perhaps no coincidence that few ethological studies of great apes, unlike those of cockroaches or salmon, have considered the species' repertoire - its "ethogram" - as a useful research tool. Finite repertoires of basic, natural units of behaviour are not obvious to observers of apes, presumably because almost all combinations of movements occasionally occur together in spontaneous behaviour. In addition, other behavioural mechanisms than trial-and-error learning and imitation may augment an individual's repertoire. For instance, play with companions or with objects may function to build up a range of routines that achieve specific results. In play, the results themselves have no biological function, by definition, but once a relation between some behavioural sequence and its result is noticed, this new routine may in future be elicited in genuine problem-solving. Since every individual is liable to have had a different history of play experiences, it is probably in principle impossible to describe "repertoires" in species which learn in this way. This makes identification of "new" behaviour very difficult, and detection of imitation consequently less likely.

The good news is that, if we instead move to a higher level of organization in the hierarchy of behaviour, then "novelty" actually becomes easier to define. In lengthy sequences of behaviour, the probability of replicating a demonstrated arrangement by chance or independent invention quickly diminishes. We would argue that the important novelty in animal or human behaviour usually consists in novel re-arrangements of elements that are themselves not novel. Our principal aim in section 2.6 will be to focus on this sort of imitation as it functions within the process of hierarchy construction. Examples of imitation employed in the process of constructing novel behaviour - as opposed to the final products shown in the well co-ordinated routines of the mountain gorillas - will be seen in orangutan behaviour. As preparation, we turn to cognitive-developmental psychology for insights into the acquisition of novel, complex behaviour.

2.5 Assembling programs by relational learning

The importance of structure has long been recognized in cognitive-developmental psychology. There, following Piaget (e.g. 1937/1954), complex behaviour is seen as constructed, by combining and co-ordinating low level components (e.g. mental, perceptual, or motor schemes) into novel sequences. Such a sequence may become integrated or fused, so that it can operate as a unified "routine"; then it can in turn be used as a component, a sub-routine, in building higher level complexes (see Case, 1985; Gibson, 1990, 1993; Greenfield, 1991; Langer, 1993). Importantly, the co-ordination and integration of components into higher level routines implies flexible modification of individual components: components must be adjustable in order to co-ordinate effectively with one another.

Hierarchical processes are considered to underlay the emergence of symbolic abilities in children, and simple physical relationships between objects, like in-ness, on-ness, between-ness, or together-ness, are often found to lie at the core of this early hierarchical behaviour. Young children between about 1.5 and 5 years of age have been found capable of resolving problems by analyzing them in terms of the object-object relationships involved, and generating goals and routines which manipulate these relationships (e.g. Case, 1985). They play with "in-ness", for instance, by repeatedly putting objects in, then out of, containers. From about their third year, they generate behavioural complexes that combine and co-ordinate several relational routines together (e.g. Case, 1985; Langer, 1996; Greenfield, 1991). These co-ordinated routines show the features of hierarchically organized behaviour, making them simple examples of behavioural programs (Case, 1985; Langer, 1996). (And a number of researchers have suggested that discovering a new way to manipulate a relationship is what nonhuman primates acquire when they learn imitatively: Visalberghi & Fragaszy, 1990; Russon & Galdikas, 1995; Russon et al, in press.)

Though this approach derives from child psychology, there is reason to expect it to apply also to great apes. Great apes and humans show very similar patterns in their early cognitive development (e.g. Gibson, 1990, 1993; Greenfield, 1991; Parker & Gibson, 1991); even as juveniles, they can achieve logical and causal reasoning skills beyond sensorimotor levels (e.g. Langer, 1993, 1996; Poti & Antinucci, in prep; Spinozzi, 1993). In adult chimpanzees, Boysen (1993, 1996) has shown numerical abilities approaching that of 3-4 year old children, including counting, summation and subtraction. Matsuzawa (1994) found that some adult wild chimpanzees who used stone hammers and anvils to crack open hard nuts also added a third stone as a wedge to level their anvil rock; human children he tested did not master this strategy until they were 6-7 years old. At 3-4 years, children are already beyond the developmental threshold at which hierarchical organization of routines appears (Langer, 1996).

We argue here that young great apes do in fact structure their learnt behaviour in ways very similar to young children. Cognitive-developmental psychology points to several diagnostics of hierarchical organization in behaviour, some we have already illustrated in gorilla food preparation and used to argue for their hierarchical organization: the iterative repetition of subroutines, and the capacity to handle optional operations interrupting the main process. Others we shall see in 2.6 include self-correction of parts of routines to meet pre-defined criteria and substitution of functionally equivalent components. Child psychology points to object-object relations as one underlying basis of hierarchically organized behaviour, and we turn now to orangutans to illustrate the object-object relations which underlie the construction of novel behavioural programs in nonhuman great apes, and the use of imitation in this process.

2.6 Case study: relational learning in orangutan imitation

Rehabilitant orangutans in Tanjung Puting National Park (Central Kalimantan, Indonesia) have provided some of the most complex examples of great ape behaviour acquired, in part, by imitation (Russon, 1996; Russon & Galdikas, 1993). These orangutans copied many unusual and standardized behavioural techniques used in the camp, including techniques for siphoning fuel from a drum into a jerry can, sweeping and weeding paths, mixing ingredients for pancakes, tying up hammocks and riding in them, and washing dishes or laundry. In some cases, their goals were clearly those of the humans, for whatever reason; in others, copying the behaviour for its own sake was apparently intended. Most of the incidents showing imitation entailed organizing many individual actions in an elaborate way. In most cases, the essence of their imitation was not specific motor actions but rather the organization of sets of actions into larger programs. Many of these programs incorporated manipulations of relations between objects - like pouring liquid into a container, threading rope through a ring, untying a knot, sweeping a path with a broom, or fanning a fire with a lid (Russon & Galdikas, 1995).

To support a broad understanding of these orangutans' imitations, we looked at a larger data set of their spontaneous orangutans' manipulations of object-object relations (700 incidents, from 700 hours systematic observation in 1990 and 1991). Some were organized as integrated, higher order programs, showing: (1) flexibility in component objects and actions (e.g. orangutans could pour water into a soda bottle, kerosene into a cup, or sand into a bag; and pouring could be done holding lip, base, or handle of a cup); (2) subordination of components to higher level goals (e.g. pouring rate was modulated to control the transfer of the substance); (3) iteration to criterion (e.g. one orangutan copied the camp technique for getting water out of a floating dugout canoe, repeatedly rocking it so that the water sloshed out; she paused periodically to inspect the water levels, then resumed rocking, and stopped finally only when almost all of the water was removed); (4) self-correction (e.g. orangutans, handed a pen upside down, would rotate it as soon as they noticed it would not write); and (5) interruption management (e.g. one managed to pour insect repellent from a bottle onto her hand while simultaneously warding off her son and daughter as they tried to butt in). This suggests that orangutans had a clear functional understanding of some object-object relations. We have not yet developed an exhaustive list of the physical relations these orangutans understood, but some of the common ones they manipulated in this integrated manner include IN-OUT (e.g. absorb by immerse, soak, and squeeze, embed by implant or scrape out, contain by pour in and scoop out, and enclose by wrap-unwrap and loosen-tighten), ON-OFF (e.g. support something on a rigid or floating base, cover by putting lid on or off), TOGETHER (e.g. join, mix, gather up, pile, tie-untie), CONTACT (e.g. touch, lean, poke, wipe, chop, hammer) and THROUGH (e.g. thread, weave). Not all orangutans showed the ability to create programs manipulating all these relations; presumably differences were a function of their varied histories.

The orangutans regularly embedded simple programs for manipulating relations within larger behavioural programs, showing how an individual's repertoire of relational programs can serve as the basis for complex behaviour. The adult female Supinah, for instance, used a "pouring" program, itself imitated from humans, several times as a subroutine within her elaborate imitation of the local behavioural strategy for making fire. She poured kerosene from a large can into a cup, poured it from the cup into the original can, poured it from the cup onto a stick, and poured it from the cup onto the fire's embers; sometimes she poured with two hands, sometimes with one. Another imitated subset of her repertoire consisted of elaborations of two relational programs, scraping or rubbing one object across another, and wetting objects in liquids; she organized these as subroutines along with a variety of actions and objects into many novel programs. This part of her behavioural "kit" probably allowed her to copy more complex behavioural programs, including sharpening axe blades, re-shaping a blowgun dart, washing clothes and floors, sawing wood, sweeping paths, and painting walls (for detailed discussions, see Russon & Galdikas, 1995). Other orangutans also made use of these particular programs as subroutines. One copied a technique for removing bark from a branch by tool-assisted scraping, and another incorporated both rubbing and wetting within her reproduction of the whole ritual tooth brushing program used by (human) camp visitors.

Although imitation of relational manipulations themselves constitutes a simple form of program level imitation, some of our cases were considerably more complex. We offer three examples to suggest the heights of hierarchical complexity these orangutans achieved and the varied levels at which their imitation could occur. Our goal here is to highlight the hierarchical structure of behavioural routines in which imitation was employed rather than to repeat arguments for the presence of imitation itself, which are detailed in Russon & Galdikas (1993). The two most complex examples are also represented diagrammatically, in an effort to clarify their organization. Diagrams represent the behavioural sequence observed, along with its inferred organization and the orangutan's probable goal and behavioural strategy. For brevity - these incidents involved 15-20 minutes of continuous activity - we describe sequences at the level of manipulations of relations between objects rather than that of the individual motor actions.

Ex. 1: Supinah steals soap and laundry by canoe (See Figure 7.) The bottom line of the figure shows the behavioural sequence observed. Those relational manipulations that were identified as integrated programs for Supinah are underlined and shown in boxes to indicate their operation as behavioural units. Shown on upper levels of the figure is the organization of her behavioural sequence inferred from the description (Russon and Galdikas, 1994). In the figure, behavioural units are joined upwards by lines to the immediately superordinate program. Supinah's overall program and her goal, again inferred, appear at the topmost level of the figure. The organization of the incident is most apparent reading from the highest level (top), down.

Supinah's goal appeared to be using the soap and laundry possessed by camp staff, who were washing laundry on a floating raft just off the end of the camp dock: this is what she achieved, she had worked to achieve this goal in the past, and her behaviour made sense only with this goal in mind.

While she could directly take the goods from the staff by intimidating them (they were afraid of her), they were protected by a guard stationed on the dock to block her access. Her overall strategy to get the soap and laundry required foiling the humans (figure level 1) and this entailed two different tactics - bypassing the guard, then taking the goods from the staff (figure level 2). Bypassing the guard meant detouring around him, which meant travelling through water because the end part of the dock where Supinah lurked stood in knee-deep water. Below this part of the dock was a dugout canoe and these orangutans are well-known for cruising down the river in pilfered canoes, but this one was moored and half full of water. Supinah dealt with this situation with two more subroutines - preparing the canoe for use, then riding it past the guard to the raft (figure level 4). Preparing the canoe had two subroutines - freeing it and baling it out (figure level 5); each involved several minutes of detailed manipulation and several relational manipulations, including untying a knot and two iterative techniques for removing water from the canoe (figure bottom level). Not shown in the figure is that Supinah interrupted her canoe preparation to climb back up the side of the dock and peek over its edge towards the guard; the guard was still there and she immediately climbed back down and resumed canoe preparation. Riding the canoe required re-orienting it relative to the dock and the raft, then propelling it alongside the dock towards the raft (figure levels 5 and bottom). Taking soap and laundry from the staff was then easy: Supinah merely hopped onto the raft; staff obligingly shrieked and jumped into the water, abandoning the soap and laundry. Supinah immediately set to work washing the clothes using most of the manipulations used in the overall washing technique standardized at the camp (e.g. rub soap on wet clothes or brush with soap, scrub clothes with brush, wring wet clothes).

The overall plan appears to have been a one-time, independent concoction and motor actions were varied flexibly in accordance with immediate needs. Supinah appeared to have used imitation, however, for the relational manipulations she deployed, sometimes in organized packages. Most clearly, she copied a standard camp technique to remove water from a boat (by rocking the floating boat side to side on the water, thereby sloshing water out of the boat over its gunwales), and possibly also the component manipulations for the particular local technique of washing laundry (e.g. rub soap on wet laundry, scrub soapy laundry with a brush, rub soap on a brush then scrub wet laundry with the soapy bush, wring wet laundry, etc.), but not its organization, since she performed the components in idiosyncratic order.

Ex. 2. Weeding paths During one of the periodic bouts of cleanup around the camp, Mr. Mursiman, a long-time staff, was cleaning paths by removing weeds that had grown along their edges. The standard technique used in the camp was to slice weeds off at ground level with a hoe, then to toss the cuttings well back into the bush; additionally, Mr. Mursiman piled the cuttings into a straight row behind him along the centre of the path before disposing of them. He reported that Siswoyo, an adult female orangutan, had followed him, watched his weeding, then started weeding herself. AR arrived to find Siswoyo about 3m behind Mr. Mursiman on the same path, also removing weeds from the side of the path and likewise piling the cuttings behind her in the path. She mostly chopped roughly at the weeds with a 5m long stick, but she sometimes pulled them out by hand, and her row of cuttings was ragged rather than straight.

Siswoyo imitated Mursiman's overall weeding program (remove weeds until a section of path is clean, pile cuttings, move to a new section of weeds, then iterate this routine until path is clean), but she may have acquired individual components independently. Both Mursiman's and Siswoyo's weed removal techniques incorporated two subroutines, tool assisted weed removal and weed piling. Mursiman's weed removal technique required co-ordinating two object-object relations, hoe-chop weeds and hoe-shave ground (so as not to disturb the soil). Siswoyo's version of weed removal was less sophisticated, incorporating only one object-object relation (tool-chop weeds), and she substituted a stick for a hoe as her tool (Mursiman had the camp's only hoe). Mursiman's weed-piling technique co-ordinated two object-object relations (collect cuttings together, arrange currings in straight row). Siswoyo's version of piling again only involved manipulating a single relationship, collect cuttings together. Siswoyo's weeding activity shows program level imitation in reproducing the hierarchical organization of the whole activity; she also copied Mursiman's programs for clearing weeds and piling them, albeit poorly differentiated versions. Lower levels of imitation were notably absent: for instance, she substituted a stick for a hoe, and used pulling rather than tool-based technique for removing weeds.

Ex. 3. Fire-making This incident lasted some 20 minutes and entailed a wide range of manipulations on 12 objects of 7 types. Figure 8 diagrams the incident, using the same notation as Figure 7. The four physical elements central to Supinah's activities are indicated as H (heat), A (air), W (wood), and K (kerosene); two elements concatenated indicate a relationship existing independently of her manipulations (e.g. HW is hot wood, wood already burning) and two elements hyphenated indicate a relationship that Supinah created (e.g. H-W is wood Supinah tried to heat, for instance by poking a stick into a fire).

Supinah's overall objective appeared to be making an active fire because, with skill, all the techniques she tried would have generated one. The elements available to her included embers of cooking fires, a large can of kerosene, a lid, a small plastic cup, and various sticks (some of them glowing hot). Her overall strategy seemed to be to combine heat, air, wood, and kerosene (figure top level). Supinah tried manipulating the relationships between these elements in three different patterns:

A-HW (blow air on hot wood)

K-HW (combine kerosene with hot wood)

A-W-K (blow air on wood after immersing it in kerosene)

We classified each of these three relational manipulations as integrated units, or small "programs", since she enacted each more than once, with variation. Her attempts were as follows (numbers indicate the sequential position of the relevant boxed descriptions on the bottom of Figure 8, counting from left to right):

1-3. A-HW blow on burning tip of stick (6 blows)

4. HW-K immerse hot stick in cup of kerosene (twice, changing the kerosene between attempts)

5-7. KW-HW touch kerosene-soaked stick to hot stick

8. A-W-K fan with lid over stick she had immersed in cup of kerosene

9. A-KW blow on tip of kerosene soaked stick

10-11. K-HW set cup of kerosene on burning embers

13. K-HW dip stick in cup of kerosene, drip kerosene on burning embers

14. K-HW pour kerosene on stick, drip kerosene on burning embers

15. K-HW pour kerosene directly on burning embers

It is perhaps significant that Supinah managed to combine three but not all four of the elements needed to make an active fire, perhaps reflecting a working memory constraint to the level of hierarchical complexity that can be achieved by a nonhuman great ape.

Despite failure to effectively execute the entire program of fire-making, Supinah's attempts show use of imitation in several ways. She probably imitated the overall strategy for making a fire, in the sense of "combine these four elements", although her version of this strategy was imprecise and inaccurate. Most clearly, she imitated several of the component techniques, eachof which represents a smaller program used as a subroutine, especially scooping a cup of kerosene from the can for use in starting fire, wetting wood with kerosene before attempting to light it, and fanning a would-be fire with a particular lid. There are also signs of that she made limited use of action level imitation: for instance, Supinah's fanning with a lid copied the specific motor technique used by camp cooks when they fanned with this same lid for this same purpose.

These examples have been chosen to show how orangutans employ imitation. Under other circumstances, however, orangutans like humans may well fail to show imitation: the situational context, the level of skill "gap" between model and observer (Parker, 1996; Vygotsky, 1962), and many other considerations may be influential. For instance, when tested with a device whose internal operation was opaque, but which could be set to deliver reward when a handle was moved in a particular way, no imitation of a human model was found (Call & Tomasello, 1995). Interestingly, one of the subject animals in this experiment was a home-reared orangutan, Chantek, that has often shown an ability to imitate arbitary human actions (Miles, Mitchell & Harper, 1996). Just such action level imitation would have been efficient for Call and Tomasello's opaque task; in section 3, we will argue that action level imitation may have evolved for other functions than skill-learning, which could help explain why is not easily recruited for skill acquisition.

2.7 Emulation of relationships or program level imitation?

It would be strange if chimpanzees lacked abilities found in gorillas, orangutans and humans - their closest relatives. We believe is that imitation may have been overlooked in chimpanzee studies, in the light of our distinction between and action level imitation. In particular, the "emulation" that Tomasello and his colleagues elicited in their experiments on imitation in chimpanzees (Tomasello et al., 1987; Nagell et al., 1993; Call & Tomasello, 1994) can be understood as program level imitation. Recall, they had demonstrators get out-of-reach food with a rake tool. In the first experiment, a demonstrator used a two-stage technique with a metal T-bar rake to get food against a wall, reaching the rake beyond the food then dragging the food in. In later experiments the demonstrator used a more standard rake with widely spaced tines but with the rake head on its back edge; in one condition, the demonstrator even showed subjects how to flip the rake from its tines onto its edge. Chimpanzees that observed these demonstrations did use the rake to bring the food within reach but they did not imitate important details of the technique demonstrated.

Tomasello rejected this behaviour as imitation, the observational learning of behavioural strategies, and attributed it instead to emulation, the observational learning of the result or goal that a demonstrator is seen to achieve (Tomasello, 1990, citing Wood, 1989). However, what the chimpanzees copied was not so much a result, "food-in-hand", as a usage, "rake-as-tool", apparently learning about the functional relationship between tool and food. We stress the relationship because the task was a tool task and tools are in essence relational: "tool" rightfully refers to objects only in terms of particular types of relations they enter into with other objects, when they mediate attaining external goals (Beck, 1980; Reynolds, 1982). Flexible tool use requires the ability to manipulate physical causal relations in a generalized manner (e.g. Parker & Gibson, 1977); and the ability to imitate tool-using tasks requires understanding the causal relations involved (e.g. Kohler, 1925/76; Piaget, 1937/1954; Visalberghi & Fragaszy, 1990; Visalberghi & Limongelli, 1996).

If great ape subjects learned something about the-rake-as-tool by watching, what they learned was de facto relational; using what has been learned about the rake-food relationship necessarily entails translating this relational learning into a behavioural strategy to bring about the goal, food-within-reach, by manipulating the relationship between tool and food. And the behavioural strategy that the chimpanzees used was basically the same as the strategy that was demonstrated. Whereas great apes trying to solve rake problems independently have been observed to use throwing, tapping and poking among other idiosyncratic relational manipulations (Parker, 1969), here they specifically enacted raking, and they raked with the head end rather than the handle. Thus their raking behaviour does match the behavioural strategy demonstrated, but at the program level rather than the action level. Chimpanzee performance on these tasks may be better captured by our concept of program level imitation than it is by the concept of emulation. We do not doubt that chimpanzees sometimes learn the affordances of objects by observation (although we do not see how this learning is cognitively less complex than imitation); however, in this case it seems an unlikely explanation. What we would have to assume is that the chimpanzees already knew about raking (the relational manipulation), but did not know that a stick is a suitable rake (the tool), and this is what they learnt by watching successful performance. No evidence of this prior ignorance is presented, and it seems improbable - what else might they have been used to raking with, if not sticks?

Examination of the techniques chimpanzees used shows that their programs of behaviour were hierarchically constructed. Their raking in the first experiment was described as simple "sweeping" motions with the T-bar rake (Tomasello et al., 1987). To be effective, however, even sweeping must operate as a relational routine, not a simple action, because it must establish a relationship of HOOKING behind the food, then maintain the hooked food-rake relationship throughout the sweeping arc of movement; in the process, the position of the food changes continuously, so SWEEPING must be modulated or corrected to track this effectively. In the other two experiments with a standard rake, effective raking requires positioning the rake head beyond the food then DRAGGING it in (and "beyond" is itself a relational state). Once correctly positioned, the rake's position relative to the food must be continually modulated to effect dragging. Several chimpanzee subjects succeeded in obtaining the food by dragging with the rake's tines rather than its edge, a very touchy relational balancing act requiring: exquisite modulation to keep the food from slipping out between the tines; when food did slip, as it usually did (Call, pers. comm.; Nagell et al., 1993), corrective measures must have been taken. The simplest correction is iterative, repositioning the rake beyond the food then resuming dragging. These characteristics - incorporating integrated relational routines as subroutines, self-correction, iteration to pre-defined criteria, and substitution of functionally equivalent components (e.g. rake flipped over, or on its tines) - all show that the chimpanzees were using hierarchically structured behavioural strategies when they manipulated the rake-food relationship, not simple actions.

We accordingly conclude that these chimpanzees did imitate the program demonstrated - although they did so at a coarse level of description. That they did not mimic some of the details demonstrated is not necessarily a sign of general cognitive weakness, but shows that program level imitation begins at higher hierarchical levels. And since the chimpanzees often succeeded, the pressure to overhaul their strategies by copying demonstrated details may have been absent under the artificial conditions of captivity and the constraints imposed by experimentation. (Compare the case of mountain gorillas, huge animals existing only on plant nutrients in a cold environment: they have considerably more to lose from retaining inefficient routines.) We conclude that chimpanzees, even without home-rearing or experience with sign languages, probably imitate in similar ways to orangutans and gorillas. It is because this imitation is typically at program level whereas researchers have sought evidence of action level imitation, and because the frequency and extent to which chimpanzees imitate at this lower level is very limited (see Custance and Bard 1994), their imitative capacity has seemed equivocal.

3.0 Discussion and Conclusions

We began this assessment of imitation in nonhumans in the conventional way, attempting to sort the wheat of cognitively complex behaviour from the chaff that can be parsimoniously explained by simpler mechanisms. We argued that most of the cases currently claimed to be animal imitation should be rejected in favour of one of these simpler explanations, response facilitation; and that, conversely, some great ape copying that has been discounted as emulation may warrant re-evaluation as imitation. With these restrictions, we concluded that: the several of the "simple" processes that guide social learning (stimulus enhancement, response facilitation, and goal emulation) can be computationally described by a single mechanism, priming of brain records. Aside from the obvious parsimony, this should help to highlight genuine cognitive complexity where it is found.

We then attempted to show that great ape imitation is hierarchically organized, using evidence from gorillas, orangutans and chimpanzees. When behaviour is viewed hierarchically, imitation is a high-level constructional ability; it is not a "special faculty" but one expression of the more general ability for constructing hierarchical cognitive structures. Thus imitation is generally found embedded within the whole process of constructing novel behavioural strategies, at various hierarchical levels. Equally, imitation will not typically be isolated from simpler processes, such as instrumental or associative mechanisms, but occur in conjunction with them. Failure to recognize these facts has led to failure to recognize imitation in great apes.

Hierarchical organization is pervasive in the nervous system, and has long been believed to apply to the coding of behaviour. Using the skilled food-gathering techniques of mountain gorillas as an illustrative example, we argued that - more than this - at least in the great apes, behaviour is organized hierarchically not simply at genetic and physiological levels, but in a way that is available to learning mechanisms under voluntary control. The hierarchical structure is made up of integrated complexes of elements, including relational combinations, generated recursively. Cognitive developmental psychologists have informed the analysis of hierarchically structured voluntary behaviour in humans; they too have found the application of such analyses to nonhuman primates to be fruitful (Gibson, 1993; Langer, 1993, 1996; Parker and Gibson, 1990; and Mitchell, 1987, whose description of "Stage 4 imitation" has elements in common with our analysis).

This view of hierarchically structured animal behaviour under voluntary control has implications for imitation. In particular, it suggests that imitation can in principle occur at many levels. For heuristic purposes, we distinguish ACTION LEVEL IMITATION (imitation of basic elements of behaviour, singly or in sequential strings) and PROGRAM LEVEL IMITATION (imitation of the organizational structure at any higher level, from single relational manipulations to the overarching program). At the program level, the matching that indicates imitation will be found not in motor action details but in their arrangement within functional programs. These can range from the program representing the overall strategy for behaviour, to programs representing any of its constituents. However, even action level imitation cannot involve copying of single muscle twitches: motor organization must be copied. Potentially, then, action level imitation might be interpreted as one end of a continuum with program level imitation at the other, varying only in the level at which organization is copied. Be that as it may, the evidence of imitation in humans (see below) and great apes fits rather neatly into the two discrete categories, with little sign of intermediates at present. The distinction may be more than heuristic, and two distinct mechanisms may be involved.

In addition to the highly-practised routines of gorillas, essential to their survival, that show signs of earlier acquisition by program level imitation, we offered cases of complex goal directed behaviour in free-ranging captive orangutans to illustrate the use of program level and action level imitation in the service of constructing new procedures. For both species, we found suggestions of limits to the hierarchical complexity that a nonhuman great ape can handle mentally. Nevertheless, our interpretation is that mountain gorillas and orangutans can imitate at the program level. Finding it hard to believe that chimpanzees should lack an ability found in their close relatives, we re-examined data previously interpreted as a sort of emulation, in which knowledge of relationships is acquired by observation. We argue that these data are better seen as signs of program level imitation, albeit at coarse levels of detail and lacking complexity, and suggest, therefore, that all great apes - not simply humans - can and do imitate at program level. While none of our data are by any means perfect, they are the best that is presently available; our current research is aimed at gaining developmental perspective on the gorilla and orangutan behaviour we describe.

While it is likely that nonhuman great apes also can imitate at action level, the importance of this for survival is more questionable. Great apes have repeatedly shown us that fine details of motor behaviour can be efficiently acquired by trial and error learning. Action level imitation of these details from other individuals, while producing convincing mimicry, may be an inefficient way of acquiring new abilities. By contrast, inefficient program organization would be tedious and sometimes impossible to "debug" by trial and error, and the organizational structure of behaviour can often be seen readily in its intermediate steps. Program level imitation makes ergonomic sense. From this perspective, it follows that imitation serving the acquisition of instrumental behaviour does not commonly operate independently, in isolation from other learning processes - indeed it may not be designed to do so. This type of imitation builds upon existing behaviour structures, and it relies on other processes like individual learning for efficiency and attunement to specific environmental contingencies. Perhaps we should re-examine how we formulate our search for "pure" imitation, imitation unadulterated by other learning processes - a conundrum that has frustrated empirical research on imitation for almost a century.

We believe that program level imitation is the major contributor to the acquisition of skilled instrumental behaviour even in humans, and that action level imitation plays a minor role. When children learn imitatively to tie shoelaces, play elaborate games, write and draw, we suspect that they seldom copy idiosyncratic details of their teachers' actions, because they more often fill in such detail by individual learning. What they imitate is the efficient hierarchical organization of actions, including bimanual co-ordinations and subroutine structure. Children's vocal imitation in acquiring speech might seem a flagrant exception to this generalization. However, recall that that the supralaryngeal tract of a young child is much smaller and less mature than that of the adult whose words are imitated; the available frequency range and articulatory capacities are quite different, and indeed the frequencies of vowel formants and phonemic pronunciation are not replicated. Instead of physical matching of sounds, as achieved so spectacularly by myna birds, the entire vowel register is shifted to a region of higher pitch. Children imitate spoken words at program level, copying the organization of phonemes into words, not the physical sounds. Action level imitation might contribute to skill learning, but perhaps more commonly as a way to retain poorly understood demonstrations in memory or to elicit further social exchange with the demonstrator than as a way to learn new instrumental behaviour (e.g. Abravanel, 1991; Moerk, 1989; Russon, 1996).

Completing our argument requires accounting for why action level imitation has been so misleadingly accepted as the prototype of imitation. There is little doubt that action level imitation does occur, even in great apes (Custance and Bard, 1994; Hayes and Hayes, 1952; Miles et al., 1996; Nagell et al, 1993; Tomasello et al 1993) and exact behavioural copying is a prominent feature in human development. (Though we suspect that action level imitation is less common in children than it seems, and often children's "imitation" may reflect response facilitation.) Many developmental psychologists have argued that action level imitation serves a social function for children - facilitating, for instance, the meshing of mother-infant behaviour and attachment, or taking on of desired social roles by "impersonation" (Nadel, 1986; Meltzoff and Gopnik 1993, Mitchell, 1987; Russon & Galdikas, 1995; Uzgiris, 1981; Yando, Seitz & Zigler, 1978). And it has been pointed out that imitation can contribute to efficient social functioning, because standardization of form is valuable in communication (Boyd and Richerson, 1988).

Perhaps, then, in great apes and humans alike, the main function of action level imitation is social. The impersonation of others' behaviour may be funny, or flattering, or an entry into a new societal role; but in all these cases, it is the look of the thing, not its effectiveness that matters. Yando et al (1978) proposed dual functions for human imitation, and our work with great apes leads to a similar conclusion. If this interpretation is correct, action level and program level are not simply prominent modes in a continuum of levels of imitative copying, but are independent processes which have evolved in response to very different needs, and thus have a very different pattern of occurrence. This would be consistent with suggestions that they are subserved by very different mechanisms: kinaesthetic-visual matching for action level imitation, but hierarchical plan construction for program level imitation.

Our "hierarchical approach" to imitation accounts for the existing results on great ape imitation better than alternative models, we believe, and helps explain the wildly discrepant views of those (mostly fieldworkers), who are sure that apes can imitate, and those (mostly laboratory workers) convinced that they cannot. In addition, it suggests new approaches to experimentation and to analysing subjects' responses:

1. Tasks should be constructed differently:

(a) If program level imitation is a process adapted to aid acquisition of complex, novel structures of behaviour, then it can only be studied with tasks having significant organizational components, a "program" that is worth copying. The logical relationships of components in the demonstrated routine need to be visible to the subjects, unlike the case in Call and Tomasello (1995), or else the task is reduced to an assay of "meaningless", action level imitation. So far, the closest experimental approach to this requirement is a sequence of two different actions, incorporated into the design of an "artificial fruit" (Whiten & Custance, 1996, p.308), however, in fact no sequence was demonstrated to chimpanzee subjects, and the "organization" of two sequential actions is not sufficiently complex to make for unambiguous analysis of copying.

(b) If action level imitation is adapted to social function, special care must be devoted to the social circumstances that will evoke imitation.

(c) If both processes can sometimes be recruited to the same learning task, a view already foreshadowed by cognitive developmental research, appropriate tasks need to incorporate complexity at several levels.

(d) If program level imitation operates in conjunction with mental apparatus capable of allowing hierarchical programs to be assembled, and with individual trial and error learning, then it would be unreasonable to expect skills to be acquired entirely by imitation; imitation will most likely be invoked in the face of difficulty, and to short-cut combinatorial complexity of possible sequences, not for trivial problems.

2. Analyses of potentially imitative behaviour need to partition the variance into hierarchical levels. Rates of copying actions, relational manipulations, and overall task-organization should be worked out separately. At the very least, reproduction of organizational structure must be separated from reproduction of motor details.

References

Abravanel, E. (1991) Does immediate imitation influence long-term memory for observed actions? Journal of Experimental Child Psychology 51:235-244.

Abravanel, E., & Gingold, H. (1985) Learning via observation during the second year of life. Developmental Psychology, 21, 614-623.

Anderson, J. R. (in press) In: Reaching into thought: the minds of the great apes, ed. A E Russon, K A Bard and S T Parker. Cambridge University Press.

Anisfeld, M. (1991) Neonatal imitation. Developmental Review 11(1):60-97.

Baddeley, A. (1990) Human Memory: Theory and Practice. Hillsdale NY: Lawrence Erlbaum Associates.

Bandura, A. (1986) Social foundations of thought and action: a social cognitive theory. Englewood Cliffs, NJ: Prentice Hall.

Bauer, P., and Mandler, J. (1989) One thing follows another: effects of temporal structure on 1- and 2-year-olds' recall of events. Developmental Psychology 25:197-206.

Beck, B. (1980) Animal Tool Behavior. New York: Garland STPM Press.

Boyd, R. & Richerson, P. J. (1988) An evolutionary model of social learning: the effects of spatial and temporal variation. In: Social learning: Psychological and biological approaches, ed. T. R. Zentall and B. G. Galef, pp. 29-48. Hillsdale, NJ: Lawrence Erlbaum Associates.

Boysen, S.T. (1993) Counting in chimpanzees: Nonhuman principles and emergent properties of number. In: The Development of Numerical Competence: Animal and Human Models, ed. S.T. Boysen & E.J. Capaldi, pp. 39-59. Hillsdale, NJ: Lawrence Erlbaum Associates.

Boysen, S.T. (1996) "More is less": The elicitation of rule-governed resource distribution in chimpanzees. In: Reaching into Thought: The Minds of the Great Apes, ed. A.E. Russon, K.A. Bard & S.T. Parker. Cambridge University Press.

Byrne, R. W. (1977) Planning meals: problem-solving on a real data-base. Cognition 5:287-232.

Byrne, R W (1993) Hierarchical levels of imitation. Commentary on M Tomasello, A C Kruger and H H Ratner "Cultural learning." Behavioral and Brain Sciences 16:516-517.

Byrne, R. W. (1994) The evolution of intelligence. In: Behaviour and Evolution ed. P. J. B. Slater & T. R. Halliday, pp. 223-264. Cambridge University Press.

Byrne, R. W. (1995) The thinking ape: Evolutionary origins of intelligence. Oxford University Press.

Byrne, R. W. (1995) Primate cognition: comparing problems and skills. American Journal of Primatology 37:127-141.

Byrne, R. W. & Byrne, J. M. E. (1991) Hand preferences in the skilled gathering tasks of mountain gorillas (Gorilla g. beringei). Cortex 27:521-546.

Byrne, R. W. & Byrne, J. M. E. (1993) Variability and standardization in the complex leaf-gathering tasks of mountain gorillas (Gorilla g. beringei). American Journal of Primatology 31:241-261.

Byrne, R. W. & Tomasello, M. (1995) Do rats ape? Animal Behaviour, in press.

Call, J. & Tomasello, M. (1994) The social learning of tool use by orangutans (Pongo pygmaeus). Human Evolution 9:297-313.

Call, J. & Tomasello, M. (1995) The use of social information in the problem-solving of orangutans (Pongo pygmaeus) and human children (Homo sapiens). Journal of Comparative Psychology 109:308-320.

Case, R. (1985) Intellectual Development: Birth to Adulthood. New York: Academic Press.

Chomsky, N. (1957) Syntactic structures. The Hague: Mouton.

Custance, D. and Bard, K. A. (1994) The comparative and developmental study of self-recognition and imitation. In: Self-awareness in animals and humans. Developmental perspectives, ed. S. T. Parker, R. W. Mitchell & M. L. Boccia, pp.207-26. Cambridge University Press.

Custance, D. M., Whiten, A., & Bard, K. A. (1995) Can young chimpanzees (Pan troglodytes) imitate arbitrary actions? Hayes & Hayes (1952) revisited. Behaviour 132:11-12.

Dawkins, R. (1976) Hierarchical organization: a candidate principle for ehtology. In: Growing points in ethology, ed. P P G Bateson and R A Hinde. Cambridge University Press.

Dawson, B. V. & Foss, B. M. (1965) Observational learning in budgerigars. Animal Behaviour, 13:470-74.

Galef, B. G. (1988) Imitation in animals: History, definition and interpretation of data from psychological laboratory. In: Comparative social learning, ed. T. Zentall & B. G. Galef Jr. Hillsdale, N. J: Erlbaum.

Galef, B.G. Jr. (1992) The question of animal culture. Human Nature 3:157-178.

Galef, B. G., Manzig, L. A., & Field, R. M. (1986) Imitation learning in budgerigars: Dawson and Foss (1965) revisited. Behavioural Processes 13:191-202.

Gibson, K.R. (1990) New perspectives on instincts and intelligence: brain size and the emergence of hierarchical mental construction skills. In: "Language" and Intelligence in Monkeys and Apes, ed. S. T. Parker & K. R. Gibson, pp.97-128. Cambridge University Press.

Gibson, K.R. (1993) Animal minds, human minds. In: Tools, language, and cognition, ed. K.R. Gibson & T. Ingold, pp. 3-19. Cambridge University Press.

Greenfield, P. (1991) Language, tools and the brain: the ontogeny and phylogeny of hierarchically organized sequential behavior. Behavioral and Brain Sciences, 14:531-95.

Guillaume, P. (1926) Imitation in children. Chicago IL: University of Chicageo Press.

Hayes, K. J. and Hayes, C. (1952) Imitation in a home-raised chimpanzee. Journal of Comparative Physiological Psychology 45:450-459.

Heyes, C. M. (1993). Imitation, culture and cognition. Animal Behaviour 46:999-1010.

Heyes, C. M., & Dawson, G. R. (1990) A demonstration of observational learning in rats using a bidirectional control. Quarterly Journal of Experimental Psychology 42B:59-71.

Heyes, C. M., Dawson, G. R. & Nokes, T. (1992). Imitation in rats: initial responding and transfer evidence from a bidirectional control procedure. Quarterly Journal of Experimental Psychology. Section B: Comparative and Physiological Psychology 45B:229-240.

Kohler, W. (1925/1976) The Mentality of Apes. New York: Liveright.

Langer, J. (1993) Comparative cognitive development. In: Tools, language, and cognition, ed. K.R. Gibson & T. Ingold. Cambridge University Press.

Langer, J. (1996) Heterochrony and the evolution of primate cognitive development. In: Reaching into Thought: The Minds of the Great Apes, ed. A.E. Russon, K.A. Bard & S.T. Parker. Cambridge University Press.

Lashley, K. S. (1951) The problem of serial order in behavior. In: Cerebral mechanisms in behavior: The Hixon symposium, ed. L. A. Jeffress. New York: Wiley.

Mackintosh, N. J. (1994) Classical and operant conditioning. In: Learning and Skills, ed. N.J. Mackintosh & A.M. Colman, pp.1-18. London: Longman Essential Psychology.

Matsuzawa, T. (1994) Field experiments on use of stone tools by chimpanzees in the wild. In: Chimpanzee Cultures, ed. R.W. Wrangham, W.C. McGrew, F.B.M. de Waal & P.G. Heltne, pp. 351-70. Cambridge, MA: Harvard University Press.

Meltzoff, A. N. & Moore, M. K. (1977) Imitation of facial and manual gestures by human neonates. Science 198:75-8

Meltzoff, A. N. & Moore, M. K. (1983) Newborn infants imitate adult facial gestures. Child Development 54:702-9.

Meltzoff, A. and Gopnik, A. (1993) The role of imitation in understanding persons and developing a theory of mind. In: Understanding other minds. Perspectives from autism, ed. S. Baron-Cohen, H. Tager-Flusberg & D. J. Cohen. Oxford University Press

Miles, H. L., Mitchell, R. W., & Harper, S. (1996) Simon says: the development of imitation in an enculturated orangutan. In: (Eds.), Reaching into thought: the minds of the great apes, ed A. E. Russon, K. A. Bard, & S. T. Parker (pp. 278-299). Cambridge University Press.

Miller, G. A., Galanter, E. and Pribram, K. (1960) Plans and the structure of behavior. New York: Holt, Rinehart and Winston.

Mitchell, R.W. (1987) A comparative developmental approach to understanding imitation. Perspectives in Ethology 7:183-215.

Mitchell, R.W. (1993) Mental models of mirror-self-recognition: two theories. New Ideas in Psychology 11:295-325.

Moerk, E.L. (1989) The fuzzy set called "imitations". In: The Many Faces of Imitation in Language Learning ed. G.E. Speidel & K.E. Nelson, pp.277-303. New York: Springer-Verlag.

Moore, B.R. (1992) Avian movement imitation and a new form of mimicry: tracing the evolution of a complex form of learning. Behaviour 122:231-263.

Morgan, C.L. (1900) Animal Behaviour. London: Edward Arnold.

Nadel, J. (1986) Imitation et Communication entre Jeunes Enfants. Paris: PUF.

Nagell, K., Olguin, R.S. & Tomasello, M. (1993). Processes of social learning in the tool use of chimpanzees (Pan troglodytes) and human children (Homo sapiens). Journal of Comparative Psychology 107:174-186.

Newell, A., J.C. Shaw, & Simon, H.A. (1958) Elements of a theory of human problem solving. Psychological Review 65:151-66.

Newell, A. & Simon, H.A. (1972) Human problem solving. New York: Prentice-Hall.

Parker, C. (1969) Responsiveness, manipulation and implementation behavior in chimpanzees, gorillas and orang-utans. Proceedings of the Second International Congress of Primatology, 1, 160-166. New York: S Karger.

Parker, S. T. (1996) Apprenticeship in tool-mediated extractive foraging: The origins of imitation, teaching and self-awareness in great apes. In: Reaching into Thought, ed. A. Russon, K. Bard, & S. T. Parker (pp. 348-370). Cambridge University Press.

Parker, S.T. & Gibson, K.R. (1977) Object manipulation, tool use, and sensorimotor intelligence as feeding adaptations in cebus monkeys and great apes. Journal of Human Evolution 6:623-41.

Parker, S.T. & Gibson, K.R. (1991) "Language" and Intelligence in Monkeys and Apes. Cambridge University Press.

Piaget, J. (1937/1954) The Construction of Reality in the Child. New York: Basic Books.

Piaget, J. (1945/1962) Play, Dreams, and Imitation in Childhood. New York: Norton.

Poti, P. & Antinucci, F. (in prep). Chimpanzee's logical development: main results.

Rescora, R. A. (1991) Associative relations in instrumental learning. Quarterly Journal of Experimental Psychology 43B:1-24.

Reynolds, P.C. (1982) The primate constructional system: The theory and description of instrumental tool use in humans and chimpanzees. In: The Analysis of Action, ed. M. Van Cranach & R: Hass, pp. 243-385. Cambridge University Press.

Romanes, G.J. (1884) Mental Evolution in Animals. New York: AMS Press.

Romanes, G.J. (1889) Mental Evolution in Man. New York: Appleton.

Russon, A.E. (1996) An everyday view of imitation: Identifying and characterizing true imitation in rehabilitant orangutans (Pongo pygmaeus). In: Reaching into Thought: The Minds of the Great Apes, ed. A.E. Russon, K.A. Bard & S.T. Parker. Cambridge University Press.

Russon, A.E., Galdikas, B.M.F. (1993) Imitation in free-ranging rehabilitant orangutans. Journal of Comparative Psychology, 107:147-161.

Russon, A.E. & Galdikas, B.M.F. (1995) Constraints on great apes' imitation: Model and action selectivity in rehabilitant orangutan (Pongo pygmaeus) imitation. Journal of Comparative Psychology, 109, 5-17.

Russon, A. E., Michell, R. W., Lefebvre, B., & Abravanel, E. (in press). The comparative evolution of imitation. In: Piaget, evolution, and development, eds.J. Langer & M. Killen. Hillsdale NJ: Lawrence Erlbaum Associates.

Spence, K. W. (1937) Experimental studies of learning and higher mental processes in infra-human primates. Psychological Bulletin, 34:806-850.

Spinozzi, G. (1993) The development of spontaneous classificatory behavior in chimpanzees (Pan troglodytes). Journal of Comparative Psychology, 107:193-200.

Thorndike, E.L. (1898) Animal intelligence: an experimental study of the associative process in animals. Psychological Review Monograph 2(8):551-553.

Thorpe, W. H. (1956). Learning and instinct in animals. London: Methuen.

Tomasello, M. (1990) Cultural transmission in the tool use and communicatory signaling of chimpanzees? In: "Language" and Intelligence in Monkeys and Apes, ed. S. T. Parker & K. R. Gibson, pp. 274-311. Cambridge University Press.

Tomasello, M., Davis-Dasilva, M, Camak, L. & Bard, K. (1987) Observational learning of tool use by young chimpanzees. Human Evolution, 2:175-185.

Tomasello, M., Savage-Rumbaugh, E. S., & Kruger, A. C. (1993) Imitative learning of actions on objects by children, chimpanzees, and enculturated chimpanzees. Child Development 64:1688-1705.

Uzgiris, I.C. (1981) Two functions of imitation during infancy. International Journal of Behavioral Development 4:1-12.

Visalberghi, E. & Fragaszy, D. (1990) Do monkeys ape? In: "Language" and intelligence in monkeys and apes, ed. S. T. Parker & K. R. Gibson, pp. 247-73. Cambridge University Press.

Visalberghi, E. & Limongelli, L. (1996) Acting and understanding: Tool use revisited through the minds of capuchin monkeys. In A.E. Russon, K.A. Bard & S.T. Parker (eds.), Reaching into Thought: The Minds of the Great Apes. Cambridge, U.K.: Cambridge University Press.

Vygotsky, L. (1962) Thought and language. Cambrige MA: MIT Press.

Waterman, P. G., Choo, G. M., Vedder, A. L., & Watts, D. (1983) Digestibility, digestion-inhibitors and nutrients of herbaceous foliage and green stems from an African montane flora and comparison with other tropical flora. Oecologia 60:244-249.

Watts, D. P. (1984) Composition and variability of mountain gorilla diets in the central Virungas. American Journal of Primatology 7:323-356.

Whiten, A., & Custance, D. (1996) Studies of imitation in chimpanzees and children. In: Social Learning in Animals: The Roots of Culture, eds. Heyes, C. M. and Galef, B. G. (pp. 291-318). San Diego: Academic Press.

Whiten, A., Custance, D., Gomez, J-C., Teixidor, P. and Bard, K. A. (1996) Imitative learning of artificial fruit processing in children (Homo sapiens) and chimpanzees (Pan troglodytes). Journal of Comparative Psychology 110:3-14.

Whiten, A., & Ham, R. (1992) On the nature and evolution of imitation in the animal kingdom: reappraisal of a century of research. In: Advances in the study of behavior, Vol. 21, Ed. P. J. B. Slater, J. S. Rosenbaltt, C. Beer, & M. Milinski, p.239-283. New York: Academic Press.

Wood, D. (1989) Social interaction as tutoring. In: Interaction in human development, ed. M. H. Bornstein and J. S. Bruner, pp. 59-80. Hillsdale NJ: Lawrence Erlbaum Associates.

Yando, R., Seitz, V., Zigler, E. (1978) Imitation: a developmental perspective. Hillsdale, NJ: Lawrence Erlbaum.

Zentall, T. R. (1996). An analysis of imitative learning in animals. In: Social learning in animals: the roots of culture, eds. C. M. Heyes & B. G. Galef (pp. 221-243). San Diego: Academic Press.

Zentall, T. & Galef, B. G. Jr (eds.) (1988) Comparative social learning. Hillsdale, N. J: Erlbaum.

Acknowledgements

We would like to thank Bob Mitchell, Juan-Carlos Gomez, and several anonymous referees for a great deal of help in improving earlier drafts of this paper. For support of the field work from which the ideas in the paper grew, we thank the National Geographic Society, the Carnegie Trust for the Universities of Scotland and the Natural Sciences and Engineering Council of Canada; and for permission and assistance for our work with the great apes, we thank the Dian Fossey Gorilla Fund, L'Office Rwandaise du Tourisme et Parcs Nationaux, the Indonesian Forestry Department and its Nature Protection Agency (PHPA), the Indonesian Institute of Sciences (LIPI), the Faculty of Biology at the Universitas Nasional, and B. Galdikas' Orangutan Research and Conservation/Project.