A study on the optimization of basketball tactics classroom teaching strategies based on deep reinforcement learning

Based on the problems of delayed state recognition, experience-based teaching of the tactics in basketball classroom, and the difficulty of continuously updating learning feedback, this paper proposes a method to optimize the teaching strategies of basketball tactics classrooms based on deep reinforcement learning. The proposed model focuses on how classroom tactical states are transformed into instructional decisions. It integrates student trajectories, ball possession, tactical roles, and teacher feedback into a unified state representation, selects intervention actions through DQN, and updates strategies according to feedback rewards through prioritized experience replay. The model integrates the positions of 10 students and the ball, ball-possession status, tactical roles, instructor intervention labels, and feedback rewards into a unified classroom state vector. It then constructs a discrete instructional action space consisting of timeout explanations, local corrections, tactical replays, role rotations, group scrimmages, and individual prompts. A reward function is further defined using indicators such as positioning accuracy, passing-trajectory deviation, help-defense response delay, and offensive-to-defensive transition time. During model development, these feedback signals were stored in the experience replay buffer and used to update the teaching policy; the resulting DQN parameters were subsequently fixed during the classroom evaluation. Experimental results showed that, compared with traditional instruction, DQN dynamic recommendation was associated with a shorter average intervention response latency (5.2 s vs. 18.7 s), a higher instructional-action agreement rate (89.4% vs. 71.6%), a higher comprehensive tactical understanding score (88.4% vs. 69.5%), a higher comprehensive tactical execution accuracy (87.7% vs. 66.7%), and a lower invalid intervention rate (7.9% vs. 24.3%). The results show that the model has the ability of improving the timeliness, relevance, and iteration of feedback of interventions in basketball tactical class; the model can be used as the technical route to verify, judge and improve the technical level of intelligent teaching of physical education classes.