Agent-based Models for Animal Cognition: A Proposal and Prototype Elske van der Vaart 1,2 1 Theoretical Biology University of Groningen Kerklaan 30, 9751 NN Haren The Netherlands elskevdv@ai.rug.nl Rineke Verbrugge 2 2 Artificial Intelligence University of Groningen P.O. Box 407, 9700 AK Groningen The Netherlands rineke@ai.rug.nl ABSTRACT Animal ecologists have successfully applied agent-based models to many different problems. Often, these focus on issues concerning collective behaviors, environmental interactions, or the evolution of traits. In these cases, patterns of interest can usually be investigated by constructing the appropriate multi- agent system, and then varying or evolving model parameters. In recent years, however, the study of animal behavior has increasingly expanded to include the study of animal cognition. In this field, the question is not just how or why a particular behavior is performed, but also what its ‘mental underpinnings’ are. In this paper, we argue that agent-based models are uniquely suited to explore questions concerning animal cognition, as the experimenter has direct access to agents’ internal representations, control over their evolutionary history, and a perfect record of their previous learning experience. To make this possible, a new modeling paradigm must be developed, where agents’ reasoning processes are explicitly simulated, and can evolve over time. We propose that this be done in the form of “if-then” rules, where only the form is specified, not the content. This should allow qualitatively different reasoning processes to emerge, which may be more or less “cognitive” in nature. In this paper, we illustrate the potential of such an approach with a prototype model. Agents must evolve explicit rule sets to forage for food, and to escape predators. It is shown that even in this relatively simple setup, different strategies emerge, as well as unexpected outcomes. Categories and Subject Descriptors I.2.2 [Artificial Intelligence]: Automatic Programming – program modifications; I.2.4 [Artificial Intelligence]: Knowledge Representation Formalisms and Methods – representations (procedural and rule-based); I.2.11 [Artificial Intelligence]: Distributed Artificial Intelligence – multi-agent systems; J.3 [Life and Medical Sciences]: biology and genetics. General Terms Algorithms, Experimentation, Theory Keywords agent-based models, genetic algorithms, evolution, animal cognition, theory of mind 1. INTRODUCTION Agent-based models enjoy quite some popularity in the field of animal ecology [8]. As they are uniquely suited to simulate individual decisions, virtual habitats, and many generations, they allow biologists to investigate questions concerning collective behaviors, environmental factors, and evolutionary scenarios. Verbal accounts of how a particular animal behavior arises can be implemented, then tested, to see whether a theory’s assumptions can reproduce empirical observations. Generally speaking, the theories subjected to such ‘simulation-based evaluation’ can be captured by models involving a fixed repertoire of agent behaviors, where patterns of interest emerge from the interactions between individuals, different parameter settings, or the evolution thereof. Such simulations have provided insights into many different biological systems. In recent years, however, the study of animal behavior has increasingly expanded to include the study of animal cognition. As David Premack [13] puts it, ‘...virtually every month another cognitive ability, thought to be unique to humans, is reported in an animal’. In these studies, the question is not just how or why a particular behavior is evoked, but also what its ‘mental underpinnings’ are. Can animals reason? Plan for the future? Infer mental states? These are fascinating questions, which are, at their core, about what makes humans different from animals – or not so different after all. The problem is that it is very difficult to conclusively establish the presence of any of these mental abilities in subjects that cannot talk. No matter how ‘cognitively impressive’ a particular behavior looks, there are always two alternative explanations for a successful performance: The necessary actions may be innate, or they may have been learned previously, without any understanding of why they solve the problem. Take the experiment by Hare et al. [9], where the question is whether or not chimpanzees have any concept of visual perspective, that is, whether they can reason about what others can and cannot see. In this experiment, two chimpanzees are put in competition over two pieces of food, placed in a central compartment. Both chimpanzees are familiar to each other, with an established dominance relationship. In two separate side compartments, the subjects wait to be granted access, with no view of the baiting. Once the compartment doors are opened, the subordinate can see both pieces of food, while the dominant’s view of one of the pieces is obstructed by a barrier (Figure 1.1). Figure 1.1. Hare et al.’s [9] experiment; image after [10]. Cite as: Agent-based Models for Animal Cognition: A Proposal and Prototype, E. van der Vaart and R. Verbrugge, Proc. of 7th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2008), Padgham, Parkes, Müller and Parsons (eds.), May, 12-16., 2008, Estoril, Portugal, pp. 1145-1152. Copyright © 2008, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. 1145