Value-Approximation based Deep Reinforcement Learning Techniques: An Overview

Abstract

Deep Reinforcement Learning (DRL) combines the power of Deep Leaning and Reinforcement learning, and has started gaining a lot of attraction in various domains. Also, it is empowering the artificial intelligence based agents which could surpass human-level performance even the tasks which was earlier thought to be best performed by humans only. In Industry 4.0, DRL based agents are enabling applications ranging from autonomous fleets and automatic process control to dynamic scheduling for complex production lines. In this context, the study of value-approximation based DRL techniques gains significance as these techniques are instrumental in enabling the concept of General Artificial Intelligence. Therefore, to understand the value-based DRL and to apply it optimally, we present an overview of the value-approximation based DRL techniques and explained how DRL built on Markov Decision Process and the Bellman equation can further improve the general applicability of the DQN model for different applications and achieve better results over the existing ones. We also explain the setup of a Reinforcement learning problem and describe its environment, state, reward-function and agents.