Learning to Play in a Day: Faster Deep Reinforcement Learning by Optimality Tightening
This addresses the practical issue of long training times in deep reinforcement learning for game environments, making it more feasible for real-world applications.
The paper tackled the problem of slow training in deep reinforcement learning by proposing a novel algorithm that combines deep Q-learning with constrained optimization to tighten optimality, resulting in drastically reduced training time and significant improvements in accuracy on 49 games in the Arcade Learning Environment.
We propose a novel training algorithm for reinforcement learning which combines the strength of deep Q-learning with a constrained optimization approach to tighten optimality and encourage faster reward propagation. Our novel technique makes deep reinforcement learning more practical by drastically reducing the training time. We evaluate the performance of our approach on the 49 games of the challenging Arcade Learning Environment, and report significant improvements in both training time and accuracy.