Sonic to Knuckles: Evaluations on Transfer Reinforcement Learning

摘要

Reinforcement Learning holds the potential to enable many systems with rapid, intelligent automated decision-making. However, reinforcement learning on embodied systems is a much greater challenge than the simulated environments and tasks which have been solved to date. A learner in an embodied system cannot run millions of trials or easily tolerate fatal trajectories. Therefore, the ability to train agents in simulated environments and effectively transfer their knowledge to real-world environments will be crucial, and likely an integral part of constructing future robotic systems. We perform experiments in an original transfer reinforcement learning task we constructed using the game "Sonic 3 &: Knuckles", evaluating two transfer learning techniques from the literature.