Coordination of parallel learning processes in animals and robots
Mehdi Khamassi · HAL (Le Centre pour la Communication Scientifique Directe) · 2014
This HDR manuscript presents research work at the interface between Computational Neuroscience and Cognitive Robotics aiming to better understand how animals and robots can display behavioral adaptation capabilities in their partially unknown and changing environment. Previous studies have shown that the mammalian brain combines parallel learning processes in different memory systems. During instrumental conditioning as well as navigation, this permits initial learning based on a model of the environment followed by the progressive expression of learned habits. In computational terms, this can be formalized as a progressive shift from model-based to model-free reinforcement learning. The manuscript presents :1) Proposed computational solutions for the coordination of parallel learning processes to explain animal behavior during conditioning and navigation ;2) Uses of learning models to analyze behavioral and neural correlates of learning ; 3) Implementations of neuro-inspired learning models in robots interacting with the real world. The manuscript highlights the gain of these exchanges between disciplines to further discuss the resulting research program.