Modification of Q-learning to Adapt to the Randomness of Environment

摘要

Q-learning is a typical model-free algorithm in reinforcement learning to achieve a goal by interacting with an uncertain environment. However, conventional Q-learning cannot reach convergence and even learns bad policies when the state transition and the immediate reward of the environment are randomly distributed. This paper gives a modification of the Q-learning algorithm by exploring a Monte Carlo method to settle the above problems. Furthermore, simulation experiments are performed to validate the modified Q-learning algorithm.