The construction of online game teaching intelligence platform based on DQN algorithm

Abstract

Training game intelligence through reinforcement learning (RL) and applying games to teaching and learning can stimulate students' enthusiasm for learning. Therefore, the study built an online game teaching intelligence platform based on deep RL to enhance students' interest and enthusiasm in independent learning and improve teaching effectiveness through the game model. The study showed that the research model, after incorporating the Softplus function, had a higher overall score compared to other models in Breakout game test after 50 game simulations, with the highest average score of 3.2. In the Flappy Bird game, the research model had the fastest growth rate and the highest increase in score, with a final score of over 120 and an average score of 63.1 for the game after the improvement of the activation function. When compared with the Double-Deep Q-network (D-DQN) model, the research model scored higher performance after training and the model curve performed better in convergence. The research model suffered less interference in noise amplitude and frequency, and performed better in adjusting the weights of the neural network. In the empirical analysis, the research model had excellent interference resistance and relatively smaller errors. The overall satisfaction score of students with the system reached over 90 points, effectively improving learning effectiveness as well as academic performance.