Emergent goal directed behavior from local predictions - Dr. Jean-Charles Quinton (CNRS, Clermont Ferrand, France)

This talk introduces a controller model for autonomous agents based on local predictions. Each prediction only applies to a particular context and only anticipates the dynamics of a reduced number of signals. However, given a large enough set of such representations, robust and flexible goal-directed behaviors can emerge from their interactions. The key features of the proposed model are the following:
 - Anticipations are normative, i.e. the agent can directly assess their validity by interacting with the environment. They are task-independent and can thus be learned in an unsupervised way as to efficiently cover the space of potentialities.
 - Spatiotemporal sequences can be formed between anticipations that partially assimilate the dynamics, implicitely defining means and goals.
 - Reciprocal inhibition forces an interpretation of the sensory flow and the selection of the most adequate actions.

 The flexibility of the model is reflected by a large range of applications, in both real and simulated environments: mechanical device control, navigation tasks, biological simulations, artificial Tetris player...