GDI: Rethinking What Makes Reinforcement Learning Different From Supervised Learning
Published in AAAI-22 Workshop on Reinforcement Learning in Games, 2022
Recommended citation: Jiajun Fan, Changnan Xiao, Yue Huang, "GDI: Rethinking What Makes Reinforcement Learning Different From Supervised Learning." In the proceedings of AAAI-22 Workshop on Reinforcement Learning in Games, 2022. https://arxiv.org/abs/2106.06232
Workshop version of GDI. Easing the gap between the estimated and the true data distribution is critical for RL, and supervised learning fails to do so; this paper builds on that difference: it generalizes Generalized Policy Iteration into Generalized Data Distribution Iteration, which unifies many RL algorithms as special cases, and reaches state-of-the-art results on Atari with 22 human world records in 200M training frames.