- METHOD AND APPARATUS FOR ADAPTIVE MULTI-BATCH EXPERIENCE REPLAY FOR CONTINUOUS ACTION CONTROL

摘要

An adaptive multi-batch experience replay technique for continuous action space control. In the adaptive multi-batch experience replay (AMBER) method, storing information tuples of samples generated based on the updated policy in a replay memory in multiple batches, random mini-batch Adjusting the size of) to reduce the average importance sampling specific gravity, calculating the average importance sampling specific gravity of each sample batch in the replay memory, for the replay memory, the calculated Dropping a batch having an average importance sampling specific gravity greater than a predetermined batch drop coefficient, and updating parameters by performing random mini-batch sampling based on the batch excluded from the drop, targeting the replay memory. You can.