TRANSFER LEARNING BASED ON FORBIDDEN RULE SET IN ACTOR-CRITIC METHOD

In this paper, we aim to accelerate learning processes in actor-critic method. We proposed the effective transfer learning method, which reduces training cycles by using information acquired from source tasks. The proposed method consists of two ideas, the method to select a policy to transfer, and the transfer method considering the characteristic of each actor-critic parameter set. The selection method aims to reduce redundant trial and error that are used in the selection phase and the training phase. We introduce the forbidden rule set, which are detected easily in the training phase, and concordance rate that measures an effectiveness of a source policy. The transfer method aims to merge a selected source policy to the target policy without negative transfers. it transfers only reliable action preferences and state values that implies preferred actions. We show the effectiveness of the proposed method by simple experiments. Agents found effective policies from, the database, and finished their training with less or same episodes than the original actor-critic method.