Research on reinforcement learning of the intelligent robot based on self-adaptive quantization

Abstract

The concept of the reinforcement learning comes from behavior psychology that takes behavior learning as trial and error, by which the states of the environment are mapped into corresponding actions. There is a question of how can the behaviourism be used to learn the actions in interaction with the environment in designing an intelligent robot. In the paper, the actions that the robot takes to avoid obstacles are taken as one class of behaviors and the reinforcement learning is used to realize behavior learning of obstacle avoidance. The quantization of the state space is very important in improving the robot's learning speed. The SOM neural network is adopted to get self-adaptive quantization of the state space. The self-organization characteristic of the SOM neural network makes it possible to solve the adaptation problem and is flexible in space quantization. The reinforcement learning is used to settle the robot learning of collision avoidance behavior based on quantization of the state space and satisfying results are obtained.