| Reinforcement learning is an important branch of machine learning.Multi-goal reinforcement learning problems are a subclass of reinforcement learning problems that study reinforcement learning problems with multiple training goals.This paper comprehensively discusses the problem of multi-goal reinforcement learning,and conducts research on the research status of reinforcement learning and the past improvement schemes of reinforcement learning based on multi-goal guidance.This paper proposes a multi-metric guided solution for the multi-goal reinforcement learning problem,covering both scenarios where the scene is partially visible and the scene is fully visible.The specific plan includes the determination of overall indicators for specific problems,the determination of periodic metrics,and adaptive adjustment.According to the proposed multi-metric-guided solution,the article proposes a systematic reinforcement learning based solution in the multi-goal Mario games where the scene is partially visible,and the multi-goal clustering in which the scene is fully visible.The article mainly has the following three research results:(1)The article proposes three elements of the multi-metric guidance scheme for the multi-goal reinforcement learning problem,and gives the relevant multi-metric guidance solutions.A complete set of solutions for index determination,phased metric determination and adaptive adjustment.The solution has clear ideas and high portability.(2)Based on the proposed multi-metric guidance scheme,the article proposes a solution in the multi-metric Mario games where the scene is partially visible.The scheme determines the accuracy rate as the overall index,and introduces the accuracy rate and semaphore as phased metric;combined with the Proximal Policy Optimization algorithm called PPO,it is adaptively adjusted according to the characteristics of Mario games.According to the experimental results,the proposed scheme can drive Mario to gain the ability to overcome difficulties,and has a good performance in the overall effect.(3)Based on the proposed multi-metric guidance scheme,the paper proposes the IRD method in the multi-goal clustering problem where all scenes are visible,an Internal metrics driven Reinforcement Learning based DBSCAN clustering algorithm called IRD is proposed.IRD determines the normalized mutual information NMI and the silhouette coefficient SC as the overall metrics,and adds the silhouette coefficient to the periodic metrics;IRD optimizes the action selection and parameter update method according to the characteristics of the two-dimensional data clustering problem combined with the Twin delayed Deep Deterministic Policy Gradient algorithm called TD3.According to the experimental results,compared with the latest DRL-DBSCAN method,the NMI of IRD is 1.2% higher,and the SC of IRD is 8.6%higher. |
