Team Flowers

Overall Objectives
Scientific Foundations
Application Domains
New Results
Contracts and Grants with Industry
Other Grants and Activities

Section: New Results

Developmental exploration in models of the formation of shared ontologies in a population of agents.

Participants : Frédéric Delaunay, Pierre-Yves Oudeyer.

We have studied how developmental mechanisms can impact importantly the efficiency of cultural evolution dynamics in the formation of shared lexicons. Thus, this work tries to gather two threads of research: semiotic dynamics and developmental sciences.

On the one hand, many models of language formation and self-organization were developed in the last decade, under the general theoretical umbrella of semiotic dynamics [57] . One of the best studied model is the naming game [44] , in which a population of agents builds a shared lexicon, associating words and meanings, through cultural evolution and self-organization. In these simulations, agents interact in a peer-to-peer manner, negotiating each time the word that shall be associated to a randomly chosen meaning. Experiments have shown that simple feedback mechanisms lead to the self-organization of a globally shared set of associations [44] [57] .

On the other hand, recent years have also seen the multiplication of models showing that a developmental approach to learning can be very efficient in real-world spaces. The general idea is based on the gradual control of the complexity of the task or skill to be learnt, “starting simple” and progressively becoming more complicated.This can be realized through the control of the number of degrees of freedom of the motor apparatus, the sensitivity of sensors, or the complexity of the learning situations, which are progressively and actively increased [1] [24] [34] .

Based in these two strands of research combined, we showed that the introduction of active mechanisms of complexity control at the level of individuals in naming game experiments allows to drastically improve the speed of convergence of the global population to a shared lexicon. This was presented in [13] .

Further more technical publications are under way.


Logo Inria