Non-supervised sensory-motor agents learning

Abstract

This text initially discusses the distribution ofneurons in a neural network with non-supervised learning, A proposal for creation and destruction of neurons based on the activity related to the concepts being recognized is introduced. This model, called activity schema, is used to define a sensory motor agent as a collection of activity schemata, each one representing a sensory-motor concept. The sensorial characteristic of a sensory-motor concept is represented by a neuron that acts as a recognizer of observable things (piaget calls these things observables). The motor characteristic is given by an action vector defined by the schema on the agent's environment and by a goal vector that the schema tries to reach. The activity schema permits a useful distribution of neurons in a conceptual space, creating concepts based on action and sensation, and not only on sensation. Such approach is inspired in the theory ofthe Swiss psychologist and epistemologist Jean Piaget, and intends to make explicit the account of the processes of continuous interaction between sensory-motor agents and their environments when agents are producing cognitive structures. A four-page version of this paper will be published by the Intemational Conference on Neural Networks and Genetic AIgorithms ICANNGA'95, to be held in Ales, France.