Implementation and Verification of Reinforcement Learning in the Automatic Assessment System for Physical Education Teaching Skills

Physical education (PE) is often criticized for being too one-sided, which leads to a lack of meaningful evaluation in the teaching process. To address this issue, this research proposes the development of an artificial intelligence (AI) sports teaching evaluation model based on reinforcement learning (RL) technology. Implementing RL algorithms that can dynamically adjust and optimize evaluation criteria based on teaching behaviors is designed to improve the objectivity and accuracy of PE teacher assessments. A dataset of PE teaching (PET) activities is used to train an RL model, such as the owl search algorithm with collaborative deep Q-network (OSA-CDQN), where different teaching abilities, including communication, demonstration, and student involvement, are measured. Preprocessing procedures for the dataset include data cleaning, normalization, and conversion of unprocessed sensor and video data into structured formats that can be used for analysis. Key teaching acts are identified using feature extraction approaches, such as convolution neural networks (CNNs) for movement categorization and OpenPose for human pose detection. The performance metrics are utilized to analyze the system’s performance, and the suggested RL-based evaluation is contrasted with conventional evaluation techniques. The research’s findings have attained a high accuracy ratio (98.9%), efficiency ratio (97.6%), prediction ratio (97.8%), accuracy (98.9%) recall (95.6%) F1-score (97.4%) and precision (97.8%). According to preliminary findings, the RL-based method considerably improves assessment accuracy and offers more comprehensive, empirically supported insights into the efficacy of instruction. This research indicates that RL has promising potential for automating PE classroom assessments, which could lead to improved teaching quality through continuous feedback loops.