BROWSE
Detail
Conference Paper
Human Interactive Learning with Intrinsic Reward
Cited 0 time in 
Share 
Share
- Authors
- Sung-Yun Park, Seung-Jin Hong, Sang-Kwang Lee
- Issue Date
- 2023-08
- Citation
- IEEE Conference on Games (CoG) 2023, pp.1-4
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Recently, deep reinforcement learning (RL) algorithms have been advanced, and especially, Importance Weighted Actor-Learner Architecture (IMPALA) outperformed human expert scores in some of the Atari-2600 games. However, in the Bowling of Atari games where there is a serious sparse reward problem, IMPALA has poor performance. The sparse reward problem, which arises from the requirement of a sequence of actions, is a significant challenge in reinforcement learning. To address this problem, we propose human interactive learning with intrinsic reward as a solution. Combining human interactive learning and intrinsic reward into an RL algorithm, we build a sequential action guiding system in the training agent. As a result of combining the feedback neural network and intrinsic reward, experimental results show efficient convergence to a higher score than the baseline algorithm in Bowling.
- KSP Keywords
- Deep reinforcement learning, Guiding system, Interactive Learning, Reinforcement Learning(RL), feedback neural network