Self-organizing developmental reinforcement learning

Science serie. The original publication is available on Springer’s website at www.springerlink.com. This paper presents a developmental reinforcement learning frame-work 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 learn-ing 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.