Research Of Robot Soccer Decision Strstegy Based On Reinforcement Learning

In recent years, with the gradual development and popular of the robot soccer, it hasalready become a hot issue of the robot research, at the same time, it has provided goodtechnical reserve and foundation for many fields such as industrial robot, agricultural robot,military robot and service robot, and it provides a convenient research platform formulti-robot system and robotics. Through this research platform, the research achievementof various disciplines can better compare and evaluate, and promote the development ofvarious disciplines. In robot soccer, decision system is the core of the whole system, thestand or fall of the decision strategy is the key to the success or fail of the robot soccer.Decision strategy can be divided into role distribution, action choice and actionimplementation three parts, this paper mainly studies the main offensive players andauxiliary offensive players action choice of the decision strategy, in order to let our playersbetter adapt different changes, and let the match benefits to our players.Owning to the robot soccer is ever-changing, the environment of the robot is unknown,it is difficult to solve the problem of robot soccer decision strategy with the traditionalmethod in this environment, and reinforcement learning algorithm can conduct continuouslearning and exploration under the position of unknown expert knowledge, in order torealize the mapping from state space to action space, through continuous trial and errormechanism to improve the robot learning ability and the adaptability to the environment.But using tradition reinforcement learning to choose action can only consider itself valuefunction and it cannot consider the influence of other agents to this agent, based on this,this paper studies a new multi-agent reinforcement learning algorithm, it improves originalinternal inference algorithm and combines internal inference algorithm with Bayesclassification algorithm, the improved algorithm can make the prediction effect moreaccurate.This paper puts reinforcement learning algorithm based on improved internalinference into the application of the robot soccer decision strategy, using this researched algorithm to forecast the probability of the different action of the agent at different moment,and then taking action strategy according to the probability, through experimental resultand analysis we can prove the effectiveness of this algorithm, it can increase the learningability of the agent and the adaptability to environment through learning.