Application Of Reinforcement Learning Based On Robot Soccer Simulation

The research of Robot soccer involves the artificial intelligence, robotics,computer vision and intelligent control, etc, Now, it is a hot issue in the study of robot.As a result of robot soccer is distributed, real-time, dynamic, asynchronous, it becomea standard platform in artificial intelligence. Simulation robot soccer has higherperformance and lower requirement than the real robot. The simulation platform cansimulate entity game of robot soccer. Scholars can apply many algorithms to theplatform and test the algorithms’advantages and disadvantages.Reinforcement learning doesn’t need to know about the environment.Reinforcement learning can obtain the knowledge and improve the strategy byinteracting with the environment. Reinforcement learning has these characteristics,such as doesn’t need to know the environment model, handle the noise and stochasticvariance. The big state space can be handled by using state compression, functionapproximation and so on. So there are effective applications in the robot soccer byusing reinforcement learning.This paper introduces the robot soccer research background, practicalsignificance and the simulation platform firstly. This paper analyzes the basicprinciple of reinforcement learning and study several classic algorithms ofreinforcement learning such as temporal difference algorithm, Q algorithm andSARSA algorithm secondly.At last, this paper applies the SARSA algorithm to single agent environment.When use the reinforcement learning method, we should notice state space’sdiscretization, action function design, reward function design and action selectiondesign etc. An improved SARSA learning algorithm applied to multi-agent systemmakes the agent’s learning much efficiency. Comparing with traditional SARSAalgorithm, the experimental results of improved SARSA verifies the improved SARSAalgorithm has superiority.