Research On Parallel Multi-task Reinforcement Learning Method For Incomplete Information Game

With the development of artificial intelligence technology,the practical application of artificial intelligence technology can be seen in more and more scenarios.As a research field that is closest to the real life of human beings,machine gaming has also received extensive attention from researchers.Among them,the incomplete information machine game scenarios have become the focus of research in machine games because of the complexity caused by incomplete information perception and the fact that they are more in line with the actual rules of human life.Deep reinforcement learning,as a powerful method for solving machine games,has been applied in a variety of gaming environments and has led to major breakthroughs in some fields.This dissertation mainly studies how reinforcement learning agent under incomplete information conditions can achieve excellent performance in multiple task scenarios.To address the problem of poor generalization of traditional single-task agents,a algorithm for parallel multi-task reinforcement learning is proposed,which trains agents in multiple scenarios in parallel to make them perform well in multiple environments.In order to speed up the training and obtain better training performance,this dissertation proposes an overall framework for multi-task reinforcement learning by combining an actor-learner framework and asynchronous actor-critic algorithm.The deconstruction of the data acquisition and training results in a significant increase in training performance.At the same time,the framework addresses the problem of data latency mismatch in off-policy training due to deconstruction,and proposes data correction schemes for learner and actor respectively to correct the correctness of the data learned by the model and improve the learning ability of the intelligences through data correction.Moreover,inspired by the hard shared representation of multitask learning,we proposes an optimization of multitask reinforcement learning in conjunction with auxiliary task learning.The auxiliary task enhances the performance of the agent from two aspects:task awareness and auxiliary control.On the one hand,by improving the agnet's ability to recognize the whole environments,to avoid the problem of the agent incorrectly using experience for decision making across multiple task scenarios,so that the agent can balance the experience learned across environments.On the other hand,for various important elements in reinforcement learning,design auxiliary control tasks to strengthen the ability of the agent to understand the environment.The ability of the intelligence to discriminate between elements of the environment is enhanced by having the intelligence predict future pixel changes.Through the combined learning of various auxiliary tasks and main tasks,the level of the proposed agent is further strengthened,and the framework of multi-task reinforcement learning is improved.This dissertation use a complex 3D video game scenario as an experimental validation platform to verify the effectiveness of each component of the multitasking reinforcement learning framework separately.The advantages of the proposed approach are demonstrated by comparing it to other intelligences and by applying the intelligences to untrained scenarios.