Deep reinforcement learning: Bridging learning and control in intelligent systems

Abstract

Deep Reinforcement Learning (DRL) has gained popularity as a new approach in artificial intelligence that successfully integrates representation learning through Deep Learning with decision making in Reinforcement Learning. In this work, we investigate basic concepts of DRL design, structural composition, and scope of usage in intelligent systems. More specifically, several benchmark tasks were assigned to test states and actions of DRL algorithms such as Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO). Empirical evaluations illustrate the potential of DRL in dealing with complex tasks. Sample efficiency, stability and interpretability issues of DRL are also reviewed in the paper. Suggestions for further work are concentrated on algorithm increase of robustness, training time and expense decrease as well as enhancement of implementation efficiency in practical environment.