The hedonic agent: A constructivist approach of abductive capacities

Abstract

The most important question that autonomous agents have to answer is how to remain viable in various and changing environments despite their bounded cognitive capacities. This question is thus the same as how their semiotic capacity to guess viable solutions emerges, that is abduction. The claim is that no learning can happen without a hedonic principle. That defines the hedonic level.

The hedonic level is presented as a cognitive paradigm: the hedonic agent can auto teach its hedonic and sensorimotor anticipations and also the meaningful and useful distinctions for these anticipations. That defines the possibility of the emergence of a job architecture, in a constructivist way.

A model of emergence of abductive capacities inside an architecture of jobs and inside jobs is proposed. This model takes into account both the limited cognitive capacities of the agent and its necessity to manage continuously its compromise between exploration and exploitation. The claim is that, inside its job architecture, the hedonic agent can use only forward policies because of its bounded cognitive capacities. The theory of bandit processes provides the optimality of such policies based on the index of Gittins and their pertinence for the compromise between exploration and exploitation. A new learning rule of reinforcement, the I-Learning rule, is proposed to evaluate this index.