Self-organizing developmental reinforcement learning

Abstract

This paper presents a developmental reinforcement learning framework aimed at exploring rich, complex and large sensorimotor spaces. The core of this architecture is made of a function approximator based on a Dynamic Self-Organizing Map (DSOM). The life-long online learning property of the DSOM allows us to take a developmental approach to learning a robotic task: the perception and motor skills of the robot can grow in richness and complexity during learning. This architecture is tested on a robotic task that looks simple but is still challenging for reinforcement learning. © 2012 Springer-Verlag.