Research On Optimization Methods Of Micro Mouse Based On Reinforcement Learning

In recent years,intelligent decision-making and control technology has been developed by leaps and bounds.It has greatly improved the robot's ability to deal with complex practical problems,and gradually developed into a national-level strategy.The Micro Mouse is a type of embedded mobile robot that can autonomously explore unknown mazes and sprint to the end with the shortest path.To run at high speed in the narrow and complex maze,the Micro Mouse has high requirements for control and decision-making performance,so it has become a long-term hot spot in the field of robotics research.The reinforcement learning is an important machine learning method that allows the agent to summarize laws from the environment.It improves decision-making performance by trying different behaviors.Aiming at the decision-making process of the maze exploration stage,this thesis proposes a novel Micro Mouse optimization method.The feasibility and effectiveness of reinforcement learning technology applied to Micro Mouse behavior selection are studied,and its decision-making performance is simulated and verified in practice.The main research contents of this thesis are showing as follows:(1)Aiming at the decision-making problem of Micro Mouse in maze exploration,the current research status at home and abroad is summarized.The control system of Micro Mouse and its key technologies are introduced in detail.Based on the analysis of the decision-making goals,the traditional method of decision-making process is explained,and the evaluation index of the Micro Mouse decision-making performance is proposed.Through the introduction of deep reinforcement learning decision-making technology,the research content is clarified and the theoretical foundation is laid.(2)Based on the analysis of the maze structure where the Micro Mouse is located,a targeted random maze generation method is proposed.According to the partially observable Markov decision process,an environmental model for maze exploration is built,and a formal environmental reward method is designed.Through visual field reconstruction technology,convolutional neural network is used to extract environmental features.By using the double deep Q network,a reinforcement learning decision model is constructed,and its training management method is explained in detail.To test the effect of the model,a simulation experiment was carried out.The results of the experiment show that the decision-making model has a 55% probability to beat traditional methods,indicating its effectiveness and the need for further optimization.(3)According to the behavioral decision-making characteristics of the Micro Mouse,it is optimized from three aspects by analyzing the problems of the standard decisionmaking model.By using the advantages of the long-short term memory network in integrating historical information,the structure of the decision-making model is optimized.A parallel exploration training method is proposed to improve the training management of decision-making models.Based on the flood pre-deduction method,a prediction model of Micro Mouse behavior effects is proposed,which realizes the decision-making fusion of artificial experience and reinforcement learning.Through simulation experiments,the effectiveness of the improved methods and the combined optimization effect are proved.According to the results of experiments,the combined application of the three optimization methods has the best performance,and the improved decision-making model has a 96% probability to beat traditional methods.(4)According to the development process of the Micro Mouse,an intelligent Micro Mouse development platform was designed and implemented.By using cloud server,My SQL database,hardware-in-the-loop simulation technology,C# and Python programming language,the reinforcement learning decision-making method is applied to the embedded devices.According to the test results of the actual application,the intelligent decision-making method can meet the real-time requirements of the embedded Micro Mouse,which proves the practicality of the research.There are 45 figures,15 tables and 55 references.