Utilizing Dynamics and Reward Models in Learning Strategy Fusion

Learning strategy (LS) fusion is our previous work in reinforcement learning (RL) framework. LS fusion fuses multiple LSs, such as transfer learning, for a single task of a robot. This paper introduces two LSs into LS fusion: an LS to learn a dynamics and a reward models, and an LS using a model-based RL method. Especially, we propose to use the MixFS dynamics model which is also our previous work. MixFS decomposes the dynamics model into the task specific elements and the task invariant elements. Thus, we can initialize the dynamics model of a task by transferring the one of the other task. In simulation experiments, we apply LS fusion with the new LSs to maze tasks of a small humanoid robot, where the primitive motions, crawling and turning, are also pre-learned by LS fusion. The results demonstrate that the new LSs improve the learning speed by using MixFS.