Self Adaptive Methods for Learning Rate Parameter of Q-Learning Algorithm

Machine learning methods can generally be categorized as supervised, unsupervised and reinforcement learning. One of these methods, Q learning algorithm in reinforcement learning, is an algorithm that can interact with the environment and learn from the environment and produce actions accordingly. In this study, eight different on-line methods have been proposed to determine online the value of the learning parameter in the Q learning algorithm depending on different situations. In order to test the performance of the proposed methods, these algorithms are applied to Frozen Lake and Car Pole systems and the results are compared graphically and statistically. When the obtained results are examined, Method 1 has produced better performance for Frozen Lake, which is a discrete system, while Method 7 has produced better results for the Cart Pole System, which is a continuous system.