Abstract. Artificial Intelligence was born in 1956 as the off-spring of the newly-created cognitivist paradigm of cognition. As such, it inherited a strong philosophical legacy of functionalism, dualism, and positivism. This legacy found its strongest statement some 20 years later in the physical symbol systems hypothesis, a conjecture that deeply influenced the evolution of AI in subsequent years. Recent history has seen a swing away from the functionalism of classical AI toward an alternative position that re-asserts the primacy of embodiment, development, interaction, and, more recently, emotion in cognitive systems, focussing now more than ever on enactive models of cognition. Arguably, this swing represents a true paradigm shift in our thinking. However, the philosophical foundations of these approaches — phenomenology — entail some far-reaching ontological and epistemological commitments regarding the nature of a cognitive system, its reality, and the role of its interaction with its environment. The goal of this paper is to draw out the full philosophical

This paper provides an accessible introduction to the cognitive systems paradigm of enaction and shows how it forms a practical framework for robotic systems that can develop cognitive abilities. The principal idea of enaction is that a cognitive system develops it own understanding of the world around it through its interactions with the environment.

Small World and Scale Free network properties characterize many real complex phenomena. We assume that low level connectivity with such topological properties, e.g., anatomical or functional connectivity in brains, is compulsory to achieve high level cognitive functionality, as language. The study of these network properties provides tools to approach different issues in behavior based Artificial Intelligence (AI) that usually have been ill defined, e.g., complexity and autonomy. In this paper, we propose a model in which situated agents evolve knowledge networks holding both Small World and Scale Free properties. Experimental results in the context of Pragmatic Games, elucidate some required conditions to obtain the expected network properties when performing complex learning.

In the study of systems the function of the system is often a good hint to how it works. In the following paper I would like to suggest that in studying or modeling a cognitive system our pre-knowledge of their functions should be treated carefuly. We should focus on the statistical distribution of the system's environemnt and the ways this distribution affects the behaviour and devlopment of the cognitive system. I will show an example of how such a focus changes the view of the immune system. I would also like to show how this new outlook on the study of cognitive systems could affect attempts at creating artificial cognitive systems.

Abstract not found

et de nationalité japonaise acceptée sur proposition du jury: Prof. M.-O. Hongler, président du jury Prof. D. Floreano, directeur de thèse