Interacting with Other Agents Without a-priori Knowledge: a Radical Interactionist Architecture Developing From the Ground-Up the Ability to Infer Motivations and Predict Behaviour

In the field of developmental agent learning, Radical Interactionism introduces a novel formulation of the problem of self-discovery of sensorimotor possibilities, environmental properties and the development of behaviours without external rewards. Agent architectures following this approach have proven capable of inferring the properties of space, building models of local environments and developing behaviours satisfying inborn motivational drives, without a priori knowledge. However, new challenges are expected when extending this approach to multiagent systems, where behaviours involve interactions with other agents driven by their own, unknown, decision systems. In this article, we propose an architecture based on interactionist principles, capable of inferring the motives of other agents whose motivational systems are not known a priori. The interactionist agent thus learns to predict the behaviour of other agents, improving its chances of interacting with them. We illustrate our proposal with an example of a working system implemented on an interactionist agent in a prey-predator multi-agent context.