Skip to Main content Skip to Navigation
Theses

Etude de l'émergence de facultés d'apprentissage fiables et prédictibles d'actions réflexes, à partir de modèles paramétriques soumis à des contraintes internes.

Abstract : The long term goal of our work is the settlement of reliable and predictable learning techniques} of basic behaviors in the robotics framework. This document is a starting point for this project.

As a first step, we argue that classical learning methods do not fulfill our request about reliability and predictibility. We think the key point of this issue is the way the communication between the learning system and its environment is modelled. We illustrate this point of view by giving a reinforcement learning example.

We introduce a formalized framework in which communication
is seen as an interaction, as in physics. Two kinds of forces are applied to the system: the reaction of the system is deduced, knowing the action of its environment and the fulfilment of a set of internal constraints. Learning ability becomes an emerging property of the system} which is the result of several reactions over time. All the possible evolutions of the system are deduced from the prior knowledge about the interaction (with no need of other parameters).

We apply our technique to a set of two interconnected sub-systems,
which global goal is to learn basic behaviors.

We prove that the first sub-system may possess as an emerging property (within some restrictive conditions) the abilities of reliable and predictable reinforcement learning and latent learning.

The second one is at a starting point. Its aim is to convert physical signals into what we call perceptive information. A selective process is involved in this task, in order to choose valid hypothesis about the evolution of the signal, from a set of hypothesis called memory. Internal constraints applied to the structure of the memory limit the valid sets of perceptive informations. In a simple case of memory, we show that these constraints lead to an equivalent of the Shannon's sampling theorem.
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00375023
Contributor : Frédéric Davesne <>
Submitted on : Saturday, April 11, 2009 - 1:13:02 AM
Last modification on : Monday, October 28, 2019 - 11:34:09 AM
Long-term archiving on: : Thursday, June 10, 2010 - 8:21:53 PM

Identifiers

  • HAL Id : tel-00375023, version 1

Collections

Citation

Frédéric Davesne. Etude de l'émergence de facultés d'apprentissage fiables et prédictibles d'actions réflexes, à partir de modèles paramétriques soumis à des contraintes internes.. Informatique [cs]. Université d'Evry-Val d'Essonne, 2002. Français. ⟨tel-00375023⟩

Share

Metrics

Record views

297

Files downloads

496