Natural Language-Based Human-Machine Collaborative Learning Games Algorithm Based on Deep Rein-Forcement Learning

• Published in: IEEE access Vol. 12; pp. 28818 - 28830
• Main Author: Na, Le
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Behavioral sciences; Collaboration; Computer & video games; Cooperation; Deep learning; Deep rein-forcement learning; Deep reinforcement learning; Electronic games; Federated learning; Games; hind-sight experience re-play; Human computer interaction; human-machine collaboration; Human-machine systems; learning games; Learning systems; Machine learning; Model accuracy; Natural language; natural language instructions; Natural language processing; Natural languages; Open systems; Reinforcement learning; reward structure; User experience; Visual perception
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3365500

Human-machine collaborative game agents are usually in an open environment, and they typically obtain behavioral information through environmental rewards. However, traditional agent environment exploration techniques are limited in reward-sparse environments. Deep rein-forcement learning was adopted to design an algorithm with adversarial sparse reward environment rewards and improve the exploration ability and the decision-making ability of agents in electronic game environments. First, a human-machine collaboration model was designed using natural language instructions to guide the rein-forcement learning process of agents based on the concept of reward construction. Then, a hind-sight experience re-play algorithm was introduced to optimize it, solving the reward problem of human-machine collaborative agents in a sparse reward environment. These experiments confirmed that the designed natural language reward construction model could achieve a score of 9.8 in the game and achieve 92% prediction accuracy. The model optimized through hind-sight experience re-play could achieve a maximum accuracy of 97.8% in achieving target instructions and ultimately obtain a game score of 9.9. As a result, the designed natural language human-machine collaboration model has good application performance in coefficient reward environment games and can obtain better scores.