A Knowledge Fusion Learning Method for Reinforcement Learning in Different Tasks
Kai Zhou, Haifeng Guo, Qingjie Zhang, Long Wang, Liang Xu · 2024
In order to improve the learning efficiency of reinforcement learning agents on new tasks, this paper designs a knowledge fusion learning method, enabling agents to acquire knowledge from multiple old tasks. First, the architecture of the knowledge fusion learning algorithm is constructed; Second, information that can represent task characteristics is analyzed; Finally, an algorithm is designed to project the feature information of old and new tasks to form knowledge. Simulation experiments show that the knowledge fusion learning method allows new agents to learn new tasks by utilizing information from multiple less-related source tasks.