Transfer Learning for Deep Reinforcement Learning
Keywords: transfer learning, knowledge transfer, reinforcement learning
TL;DR: The work proposes an RL-specific modification of CycleGAN, which ensures one-to-one knowledge transfer between different RL tasks
Abstract: Deep reinforcement learning (RL) has shown the potential to achieve superhuman performance in solving complex decision tasks. Although, unlike humans, it fails to generalise and reuse previously acquired knowledge effectively, which is a crucial ability for a truly intelligent agent. The work proposes an RL-specific modification of CycleGAN, which ensures one-to-one knowledge transfer between different RL tasks. We evaluate the approach on the 2-D Atari game Pong and compare it against two baselines: using GAN and CycleGAN methods. The results demonstrate that our method consistently outperforms the state-of-the-art methods.
Submission Number: 18
Loading