In this paper, for agents working on POMDP, a learning algorithm combining the memory-less learning and the memory-based learning is proposed. At first stage of the propposed algorithm, memory-less learning is applied. As a memory-less learning algorithm, the stochastic gradient method is employed. While the first stage, a state-action set series that accmplish the task is stored in memory. In the second stage, the memory-based learning is applied. In this process, only the series that obtained the first stage is used, so that this method is able to reduce the number of required memory effectively.
The proposed algorithm are applied three kinds of simulation to be compared with memory-less learning algorithm. Through the computer simulations, it shown that the proposed algorithms works effectively in POMDP than ordinary memory-less learnings.