Efficient Practice for Deep Reinforcement Learning

Deep reinforcement learning has demonstrated its ability to solve diverse hard problems, which were not able to be solved previously. However, one of the main drawbacks is the required long training time for a task. Furthermore, trial-and-error in reinforcement learning is often inappropriate in most real-world tasks. There were some discussions on transfer learning, specifically sim-to-real transfer, but the design of new environments or defining an environment with similar or relevant goals require tedious human efforts. This research aims to leverage practice approaches, which do not require a new environmental design for transfer learning, to shorten the time for training a deep reinforcement learning agent and to achieve an optimal policy for the maximum expected rewards. We verify the efficacy of existing practice approach with Deep Q networks (DQN) in Atari games and also introduce a novel practice approach involving iterations of short periods of practice and reinforcement learning, to further improve the performance of an agent.