Enhanced Reinforcement Learning with Targeted Dropout

Abstract

In modern ages, the study on Reinforcement Learning (RL) has driven on Deep Q-Network (DQN) optimization learning prediction and control of Markov decision processes (MDPs). In this paper, the researcher used the Targeted Dropout strategy for RLs DQN that makes straight into learning and would be necessary to deal with MDPs with huge or continuous state and action spaces. Every weight/unit update, the targeted dropout selects a set of elements and to keep only the weights/units of maximum amount, and then apply dropout to the set. It has also a common pruning strategy so focus on fast approximations, such as removing weights with the smallest value or ranking the weights/units by the sensitivity of the network design and even rating by the sensitivity of the task execution with respect to the weights/units and removing the least-sensitive ones. The result shows that the proposed algorithm for enhancing the RL's DQN is more accurate in finding the best action to learn to achieve maximum reward. The simulation presents that in a minimal run of episodes it can achieve the maximum average reward, while without Targeted Dropout it takes more runs to achieve the average reward, and throughout the assessment of the algorithm, the suggested algorithm acquires more learning in finding the large reward value.