Reviewed by
Itzhak Benenson
Department of Geography and Human Environment
University of Tel Aviv
Ramat-Aviv
Tel-Aviv
Israel
69978
When developing agent-based models of residential dynamics or land use changes in the city, I often apply ‘imitation’ as one form of agent behavior. I always base its use on a common sense understanding of the notion, and simply force my model agents to behave - that is, to make decisions - similar to those made either by the majority of their neighbours, by the members of their socio-economic group or by their ‘modelled friends’. Yet, although I am frequently unsure if my intuition about the concept’s meaning is sufficient to properly formalise the process, I’ve managed to delay review of the professional literature to a more convenient time. That time arrived with receipt of the 600 page Imitation in Animals and Artifacts, a collection of papers presented at the conference "Imitation in Animals and Artifacts" held in Edinburgh in 1999. Reading the book has not convinced me to dismiss my common sense views on the subject. But, I admit, it has extended my understanding about the forms of imitation, its relation to learning, current techniques for investigating imitation in humans and animals and for formalising such behaviour in robotics and Artificial Intelligence.
The editors state that: "Building robots and software agents that can imitate other artificial or human agents in an appropriate way is an endeavour that involves the deepest problems of connecting perception, experience, context and actions." Carrying that thought further, I began to ponder its implications. If imitation, as an essential component of human learning capacities, can be properly formalised, robots will become capable of learning by imitating humans or one another. The science fiction future would have arrived!
Proceeding to the 22 papers that comprise the book, those I enjoyed most were those combining experimentation and theory. Individual behaviour, like any multi-dimensional dynamic process, is registered by recording the (vector) states of the objects that participate in the process given the (vector) states of environmental objects. The crucial factor here is determination of measurable state variables and interpretable measures of likelihood for the two processes. The papers presented certainly provide a wide spectrum of these variables and measures.
Two papers by the editors (Kerstin Dautenhahn and Chrystopher L. Nehaniv) introduce the field. The first paper, "The Agent-Based Perspective on Imitation", characterises imitation behaviour in an agent-based fashion, that is, in relation to a social environment that includes other agents. The correspondence problem of relating the behaviour of the demonstrator agent to that of the imitator is formulated. After a short presentation of the modern view on weak and strong agency, the authors illuminate the notions of agent, social environment and interactions between them, as well as clarifying the differences between imitation and learning. In the subsequent paper, they discuss the correspondence problem formally and present it as the challenge of comparison between multi-dimensional dynamic processes. The discussion makes clear that success in the field can be based on experiments and only on experiments. Dautenhahn and Nehaniv support the view that when states are properly defined and estimated, the criteria for comparison of two dynamic processes developed by mathematical statistics (from the 1970s onwards) can then be applied. They refer the reader to the papers that specify sets of states for different types of imitation behaviour and situations, together with the criteria to be utilised.
The presentation of experimental results begins with a paper by Louis M. Herman, "Vocal, Social and Self-Imitation by Bottle Nosed Dolphins". Based on his study of vocal imitation among dolphins, the author distinguishes between their ability to imitate computer-generated sounds and signature whistles, and between social imitation, when either two dolphins execute the same action upon the trainer’s signal, or when there is dolphin-dolphin and dolphin-human imitation. He also considers imitation of human gestures, of television models, self-imitation and training as imitation. All the experiments demonstrate the dolphin’s impressive imitative capabilities. Next, Irene M. Pepperberg, in "Allospecific Referential Speech Acquisition in Grey Parrots (Psittacus erithacus): Evidence for Multiple Levels of Avian Vocal Imitation" studies the parrots’ imitation of human speech – an intriguing phenomenon to my mind. Parrots’ ability to generate sounds heard from humans allows for the precise analysis of the imitation process, which is very complex and hierarchical when considered in its details. The research demonstrates the fine differences crucial for description of the parrots’ vocal imitation at different levels of the proposed hierarchy.
Johannes Fritz and Kurt Kotrschal, in their paper "On Avian Imitation: Cognitive and Ethological Perspectives", consider imitation as part of the learning process. The paper reviews numerous approaches to the experimental study of imitation in birds and theoretical frameworks that might support these approaches. Not surprisingly, they conclude that speculation far exceeds what we actually know about the subject.
In "Imitation of Sequential and Hierarchical Structure in Action: Experimental Studies with Children and Chimpanzees", Andrew Whiten demonstrates, with the aid of experiments using artificial fruit and Plexiglas boxes, that chimpanzees and 3-year-old children can successfully imitate linear sequences of four actions. Children are observed to imitate more complex sequences, including hierarchical ones, illustrated by the description of one form of a successfully imitated two-level hierarchy of action.
"Do Monkeys Ape? – Ten Years Later" presents a critical review by Elisabetta Visalberghi and Dorothy Fragaszy of the progress made during the ten years since publication of their famous paper. In 1990, they viewed experiments on primates’ ability to learn by imitation and concluded that the positive opinions on the subject reflect primarily anecdotal observations and anthropomorphic interpretations. Moreover, they brought forth evidence in the original paper that monkeys did not learn novel behaviours by imitation. Nowadays, the authors limit themselves to a strict definition of imitative learning. Namely, the copy of behaviour C is imitative if: (1) C is similar to model’s behaviour M; (2) Observation of M is necessary for production of C; (3) C is designed to be similar to M; and (4) C is a novel behaviour, not yet organised in that precise way within the organism’s repertoire.
To further verify the existence of imitation learning in monkeys, the authors carried out carefully designed experiments, the results of which they discuss. They still doubt explanations of monkey behaviour as true imitative learning and propose various alternatives.
Turning to our understanding of the imitation process, Jason Noble and Peter M. Todd, in "Imitation or Something Simpler? Modelling Simple Mechanisms for Social Information Processing", consider learning in animals from an evolutionary perspective. They claim that the transmission of information between the members of society depends on the rate of environmental change. In slowly changing environments, information should be transmitted genetically, rapid changes are best tracked by individual learning, and social learning is most suited to an intermediately changing environment. They discuss the latter situation using recent agent-based simulation models based on intentionally simple rules such as social facilitation (don’t do anything unless others are nearby), contagious behaviour (if others are fleeing, flee also), stimulus enhancement (follow someone older than you and then learn from whatever happens) and so forth. These simple forms of behaviour are convenient for simulation; even if individuals only mimic the results, meaning that imitation cannot be accepted as the true explanation, the behaviour can still stimulate null hypotheses for further development.
Henry Leiberman, in "Art Imitates Life: Programming by Example as an Imitation Game" introduces "programming by example’, the goal of which is to teach computers to programme. To learn programming, the computer records human examples of (simple) sequences of algorithmic actions, thereby becoming able to create sequences itself. A human "tutor" can continue to correct the computer during execution. Research in the field is quite inchoate, with many basic questions unanswered, including whether the qualitative features of learning, such as establishing generalised procedures and capacity to adapt to changing conditions can be ingrained in this way. Despite its preliminary nature, this is definitely an interesting line of research for computer scientists.
Another implementation of imitation procedures is "Learning to Fly" by Claude Sammut, Scott Hurst, Dana Kedzier and Donald Michie. The authors begin the paper with two nice metaphors for distinguishing between static and dynamic imitation and learning. Medical diagnosis, a system that tries to construct rules of classification, eventually distinguishes quite successfully between patients’ states; these procedures are based on static data and usually do not copy temporal relationships explicitly. In contrast, flight is a dynamic process. The paper describes an attempt to imitate piloting in order to teach an autopilot to fly. The approach taken is very close to the previously described ‘‘programming by example": The computer records the actions of a human pilot in a flight simulator and later represents them by means of a simple but adequate lexicon of elementary actions. The lexicon makes it possible to generate flight rules for each pilot tested. Comparisons between the auto-flights and those made by real pilots indicate that the autopilot flies quite well, especially under problematic landing conditions involving large gusts of wind, a situation where mechanical feedback systems often fail. The interesting fact is that imitation based on actions taken by good pilots, who make many corrections during the flight, is inferior to the autopilot operations constructed on the behaviour of pilots frugal in their use of controls. The latter pilots simply provide examples of what to do when things go wrong. For these "reacting pilots", automatic procedures "clean up" the outcome, making the automatic flight smoother than the human one.
The flight environment makes it possible to formulate clearly the basic problem of modelling imitation processes. Autopilot rules are constructed with a lexicon that is too primitive for human experts. The latter think differently. The hope expressed in this paper is that the more we understand formally the ways in which humans think, the closer the lexicon will become to real language.
Robert W. Mitchell, in "Imitation as a Perceptual Process", presents experiments with different animal species and claims that perceptual matching and one specific form of the process - kinesthetic-visual matching - is essential for bodily and facial imitation and self-recognition. The author proposes that recognition of matching between and imitation of diverse perceptual modalities is the basis for the social psychological development that (in itself) seems unlikely in non-biological entities. Many of the book’s papers are devoted to implementing imitation abilities in robots. Aude Billard, in "Imitation: A Means to Enhance Learning of a Synthetic Protolanguage in Autonomous Robots", investigates the role of imitation in learning by autonomous robots that were taught a basic synthetic proto-language via DRAMA, a fully connected recurrent neural network without hidden layers. The robot grounded a lexicon concerning its perception and learned to combine lexical terms to describe various situations. Learning was unsupervised and resulted from self-organisation of the robots’ connectionist architecture. For example, the sensor for right arm moves was activated thereby enabling the sentence ‘I move right arm’ to be learnt.
The teacher agent (either robot or human) had no access to the robot’s internal state and thus did not direct its teaching in response to the robot’s performance. The teacher, however, implicitly guided learning as the robot’s movement and perception attention were co-ordinated to those of the teacher. The imitation process essentially improved the efficiency and speed of learning. This work can therefore be considered as a step towards teaching robots languages that are more complicated.
John Demiris and Gillian Hayes present a computational architecture for equipping robots with the capacity to imitate in their paper "Imitation as a Dual-Route Process Featuring Predictive and Learning Components: A Biologically Plausible Computational Model". The proposed architecture has an active and a passive route. Within the novel active route, the imitator mentally places itself in the position of the demonstrator and internally executes candidate behaviours. Selection is based on the accuracy of prediction for the demonstrator’s incoming states, perceived as the demonstration unfolds. If none of the stored behaviours predicts an outcome sufficiently well, the passive imitation route is taken, along which the imitator learns the demonstrated behaviour and adds that to its repertoire. This architecture can also be considered as a model of primate imitation mechanisms.
Cynthia Breazeal and Brian Scassellati, in "Challenges in Building Robots That Imitate People", claim that social learning in robots demands that we formulate anew the answers to self-evident questions - how does the robot know when to imitate, what to imitate, how to map observed actions into behavioural responses, evaluate its successes and correct its actions and so on. Famous Kismet and other ‘pretty looking’ robots familiar to most of us from TV, implement anthropomorphic views of imitation and provide possible answers to the associated conundrums. Kismet necessarily has skills for face and skin colour detection; its context-sensitive attention system produces expressive displays regulated by social exchanges. The authors claim that progress in implementing social learning systems in robots forces us to address those same issues as they apply to biological systems although they are not yet being investigated.
"Sensory Motor Primitives as a Basis for Imitation: Linking Perception to Action and Biology to Robotics", by Maja J. Mataric, presents a method for imitating movement based on end-points - ends of fingers and the head, among others. The observed movements are classified and mapped onto a set of motor primitives, and a learning mechanism for creating novel sequences is formulated and implemented in a series of robots displaying different degrees of freedom of movement.
In their theoretical paper "Three Sources of Information in Social Learning", Joseph Call and Malinda Carpenter propose that we distinguish between emulation, imitation and mimicry, based on actions carried out during the behaviour, and adoption of the behaviour’s goal by the behaving object. Imitation demands both copying the action and adopting the goal, mimicry means that the action is copied although the goal is neither understood nor adopted, whereas emulation understands and adopts the goal but lacks the capacity to copy the action appropriate for achieving that goal. Behaviour is in large part defined by this triple: action, goal adoption and results. This three-tiered view is proposed as a framework for describing social learning processes.
A neuronal mechanism of imitation is offered by Michael A. Arbib in his "The Mirror System, Imitation, and the Evolution of Language". He bases this mechanism on the recent discovery of a grasping region in monkey brains - the neurons of this region are active not only when the monkey executes a specific hand action but also when the monkey observes other primates carrying out the same action. This mirror system - which matches observation and execution - is a homologue of Broca’s area, a crucial speech area in humans, with the homology assumed to be a missing link supporting the long-argued hypothesis that primitive forms of communication based on gesture preceded speech in the evolution of language. The neural code for execution and observation of hand movements might thus be the precursor of a crucial language property, namely that an utterance usually relays a similar meaning for speaker and hearer. The author further refines this view by suggesting that imitation plays a critical role in human language acquisition and performance and that brain mechanisms supporting imitation were decisive for the emergence of homo sapiens.
Michael Oliphant, in "Rethinking the Language Bottleneck: Why Don’t Animals Learn to Communicate?" considers three features of human language as fundamental for communication. These are: Syntax – conveying structured meaning through the use of structured forms; Learning – passing syntax from one generation to the next via cultural transmission; and Symbolic reference – mapping between basic lexical elements and their meanings that is arbitrarily yet conventional. Generally speaking, the focus in studies distinguishing human language from other forms of communication is on syntax. Oliphant argues that two other features – learning and symbolic reference — taken together make human language unique among existing systems of communication. Other species, say, chimpanzees, have innate symbolic systems while many animal species have learning abilities, but only humans are capable of cultural transmission of arbitrary references.
In "Transformational and Associative Theories of Imitation", Cecilia Heyes surveys a range of existing theories of proximate psychological mechanisms and imitation. To name just a few, she compares the basics of Perceptual Opacity, Transformational theories, Social Cognitive theory, Active Intermodal Matching, Associative theories, Contiguity, Reinforcement, Matched Dependent Behaviour and Copying. It turns out that these theories have quite a number of characteristics in common; the author tries to reduce the overlap by outlining a new, quite constructive theory, "Associative Sequence Learning". This theory considers each action as a sequence of simpler "action units"; imitation is reproduction of the sequence. Although the performer knows each element, the sequence has yet to be performed in modelled order. Detailed aspects of the formalisation of this reproduction process are proposed.
Stefan Vogt’s "Dimensions of Imitative Perception-Action Mediation" reviews experimental observations of imitative actions. While these observations look primitive, direct and almost inescapable, their detailed description might become a key for imitating more complex actions. The author demonstrates that the parameters of the description are different for each type of imitation and for the task to be fulfilled. A distinction is proposed between imitation of the parametric details of the action and imitation of the different actions.
Harold Bekkering and Wolfgang Prinz, in "Goal Representations in Imitative Actions", propose that instead of concentrating on whether or not a specific action can be called imitation, it might be worth concentrating on when to imitate, how to imitate and what to imitate – that is, the goal of imitation may be more important than the means. The authors propose a goal-directed theory of imitation based on the suggestion that processes enabling observers, such as children, to achieve the goals of the observed action (but not necessarily the means for their achievement) mediate imitation. These goal-directed processes in turn activate a motor program for their fulfilment.
In his article "Information Replication in Culture: Three Modes for the Transmission of Culture Elements through Observed Actions", Oliver R. Goodenough proposes modes for the transmission of culture: non-linguistic transmission, stories and formulas. According to this view, all these modes of cultural transmission replicate actions rather than ideas. Goodenough proposes decoupling the transmission of language-based elements into their separate actions, which he explains against the background of human phenomena such as hypocrisy.
To conclude, ‘imitation’ (like ‘agent’ and ‘behaviour’) belongs to a set of basic notions. The experimental study of imitation in humans and animals (and experts’ view of these results) are crucial for the commonsensical model developer, regarding both the formulation of the model, interpretation of the results and relating them to the social theory. That is why Imitation in Animals and Artifacts is a very useful book. It provides a wide perspective on the field, should be recommended to researchers in the behavioural, social, computer and robotic sciences and, inevitably, to all modellers seeking to simulate real-world phenomena with the help of artificial agent societies.
Return to Contents
of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2004