Performance analysis of Dyna-Q algorithm in unstable environment
Bakun Zhu, Weigang Zhu, Wei Li, Jiaxin Li, Yang Ying · International Conference on Intelligent Equipment and Special Robots (ICIESR 2021) · 2021
Dyna-Q algorithm has excellent convergence performance and is widely used in planning, control and other fields. Most of the real environment is unstable environment, so it is of great significance to study the performance of Dyna-Q algorithm in unstable environment. In this paper, an environment with controllable instability is constructed and the influence of environmental instability on algorithm performance is studied. Experimental results show that Dyna-Q has better convergence performance than Q-Learning algorithm except in extremely unstable environment, and has the same algorithm validity as Q-Learning algorithm.