Respective Advantages and Disadvantages of Model-based and Model-free Reinforcement Learning in a Robotics Neuro-inspired Cognitive Architecture

摘要

Combining model-based and model-free reinforcement learning systems in robotic cognitive architectures appears as a promising direction to endow artificial agents with flexibility and decisional autonomy close to mammals. In particular, it could enable robots to build an internal model of the environment, plan within it in response to detected environmental changes, and avoid the cost and time of planning when the stability of the environment is recognized as enabling habit learning. However, previously proposed criteria for the coordination of these two learning systems do not scale up to the large, partial and uncertain models autonomously learned by robots. Here we precisely analyze the performances of these two systems in an asynchronous robotic simulation of a cube-pushing task requiring a permanent trade-off between speed and accuracy. We propose solutions to make learning successful in these conditions. We finally discuss possible criteria for their efficient coordination within robotic cognitive architectures.