GDR MePhy

http://centuri-livingsystems.org/pdp2018-11/

Why animals learn the way they do?
An experimentally driven computational approach

Supervisors
David Robbe / INMED / david.robbe@inserm.fr
Christophe Eloy / IRPHE / christophe.eloy@irphe.univ-mrs.fr

Abstract
By constantly interacting with their environment, animals are capable of developing adaptive strategies to maximize reward collection, avoid punishments and minimize energy expenditure. The biological algorithms underlying trial-and-error learning are largely unknown. To address this question, we will examine whether different computational models can reproduce the learning dynamics and behavioral strategy of rats in a laboratory-based task. The data to model are already acquired and come from experiments in which animals, engaged in a multi-trial time estimation task, converged progressively towards a conserved embodied strategy. Learning models will be based on classical reinforcement techniques and more recent developments coming from artificial intelligence (deep learning). In this project, the back and forth interaction between experiments and theory will further the understanding of the mechanisms underlying learning and trial-by-trial adjustment in performance.

Keywords
Adaptive behavior; Reinforcement learning; Robotics; Embodiment; Motor learning; Rats; Artificial Intelligence, Machine learning