Real-time Automated Guided vehicles scheduling with Markov Decision Process and Double Q-Learning algorithm

Abstract

As an outcome of Industry 4.0, Automated Guided Vehicles (AGVs) are becoming increasingly popular as a mode of mobility on flexible shop floors for material handling. The efficiency of such systems is determined by a variety of parameters, including collision and deadlock-free operation. In multi-AGV systems, it is challenging to plan the shortest path for each AGV without conflict or collision. However, significant hurdles remain on the real-time scheduling of AGVs due to the high dynamics, difficulty, and uncertainty of the shop floor environment. This paper proposes a Deep Reinforcement Learning (DRL) technique based AGV real-time scheduling strategy with a set of rules to overcome these obstacles. AGV real-time scheduling is first defined as a Markov Decision Process (MDP) with the representation of the state, action, reward and optimal mixed rule policy. Following that, a novel Deep Reinforcement Double Q-Learning is used to schedule different AGVs to different states. Ultimately, the suggested technique exhibited superiority over the previously developed real-time Q-learning AGV scheduling technique.