HAL Id: inria-00536870
https://hal.inria.fr/inria-00536870
Submitted on 17 Nov 2010HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Sensorimotor Exploration/Exploitation with
Coordinating Local Predictions
Jean-Charles Quinton
To cite this version:
Jean-Charles Quinton. Sensorimotor Exploration/Exploitation with Coordinating Local Predictions. 4th International Conference on Cognitive Systems (CogSys 2010), Jan 2010, Zürich, Switzerland. �inria-00536870�
Reinforcement, repetition
S
ensorimotor
E
xploration/
E
xploitation
with coordinating local predictions
Jean-Charles Quinton (
[email protected]
)
C
og
S
ys
2010
F
ramework
&
P
rinciples
This contribution aims to show how exploration and exploitation might be tightly intertwined when modeling sensorimotor behaviors with coordinated predictive local representations. In such a framework, learning equals to
creating and selecting anticipations to adapt to the dynamics of the agent and its environment. Motor actions are
undertaken based on the expected outcome of the anticipations, and anticipations reinforced when successfully matching the dynamics. Reaching goals is thus equivalent to navigating through the sensorimotor space by forming and following
chains of coordinated predictions. Although the agent may constantly only try to exploit its knowledge, the presence of
multiple dynamic goals, the lack of correct anticipations, interactional noise or external constraint will lead to further exploration and the generation of new task-independent representations.
C
ORTEX
P
roject
LORIA/INRIA, 615 rue du Jardin Botanique
54600 Villers-lès-Nancy France
I
SM
G
roup
IRIT-ENSEEIHT, 2 rue Charles Camichel
31071 Toulouse Cedex 7 France
New anticipations coordinate by propagating their activity, transform the landscape and thus lead to new exploration
Motor babbling… Imitation,
guided action…
Sensorimotor interactions are mastered and allow flexible and efficient reaching or stable goals
Time / development
Learned anticipations combine with renewed attractors to adapt the dynamics and further explore the sensorimotor space
New constraints, bodily growth, coincidental exploration
Innate reflexes and drives shape the initial attractor landscape…
Reinforcement, alternation between behaviors
Anticipations are introduced, confirmed or revised based on repeated interactions
BEGINNING
Agent’s life begins, anticipations are acquired by interacting
