Collaborative multi-agents in dynamic industrial internet of things using deep reinforcement learning

摘要

Abstract Sustainable cities are envisioned to have economic and industrial steps toward reducing pollution. Many real-world applications such as autonomous vehicles, transportation, traffic signals, and industrial automation can now be trained using deep reinforcement learning (DRL) techniques. These applications are designed to take benefit of DRL in order to improve the monitoring as well as measurements in industrial internet of things for automation identification system. The complexity of these environments means that it is more appropriate to use multi-agent systems rather than a single-agent. However, in non-stationary environments multi-agent systems can suffer from increased number of observations, limiting the scalability of algorithms. This study proposes a model to tackle the problem of scalability in DRL algorithms in transportation domain. A partition-based approach is used in the proposed model to reduce the complexity of the environment. This partition-based approach helps agents to stay in their working area. This reduces the complexity of the learning environment and the number of observations for each agent. The proposed model uses generative adversarial imitation learning and behavior cloning, combined with a proximal policy optimization algorithm, for training multiple agents in a dynamic environment. We present a comparison of PPO, soft actor-critic, and our model in reward gathering. Our simulation results show that our model outperforms SAC and PPO in cumulative reward gathering and dramatically improved training multiple agents.