Parameters Optimization for Reinforcement Learning with Nonlinear Time-Varying Strategy by Using Uniform Experiment Design

This paper optimizes reinforcement learning (RL) parameters and makes agents solve problems more accurately and efficiently. The RL algorithm that the paper used is Q-learning, and the experimental environment is mazes. There are three parameters to influence the entire performance in RL, such as learning rate, greedy factor, and discount rate. This paper introduces a nonlinear time-varying strategy (NTV) into the Q-learning algorithm and uses a uniform experiment design (UED) to effectively obtain the best parameter combination. The agent can steadily figure out a solution to the mazes with the fewest steps. In conclusion, this paper proposes a Q-learning with NTV (NTV-Q-learning) to search the solutions and steadily get out of the mazes with minimal steps to prove the effective improvement of Q-Learning. From the experimental results, the solution obtained by NTV-Q-learning is better than Q-Learning.