A deep reinforcement learning-based method applied for solving multi-agent defense and attack problems. (15th August 2021)

• Record Type: Journal Article
• Authors: Huang, Liwei; Fu, Mingsheng; Qu, Hong; Wang, Siying; Hu, Shangqian
• Is Part Of: Expert systems with applications. Volume 176(2021)
• Publication Date: 2021-08-15
• Subjects: Multi-agent cooperation -- Defense and attack -- Deep reinforcement learning -- Multi-agent reinforcement learning; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2021.114896
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: The multi-agent defense and attack environment is reconstructed. Several algorithms are applied to solve the considered problem. We redefine the state space, the action space and the reward functions accordingly. Comparison experiments are conducted to show the performance of the employed models. Abstract: Learning to cooperate among agents has always been an important research topic in artificial intelligence. Multi-agent defense and attack, one of the important issues in multi-agent cooperation, requires multiple agents in the environment to learn effective strategies to achieve their goals. Deep reinforcement learning (DRL) algorithms have natural advantages dealing with continuous control problems especially under situations with dynamic interactions, and have provided new solutions for those long-studied multi-agent cooperation problems. In this paper, we start from deep deterministic policy gradient (DDPG) algorithm and then introduce multi-agent DDPG (MADDPG) to solve the multi-agent defense and attack problem under different situations. We reconstruct the considered environment, redefine the continuous state space, continuous action space, reward functions accordingly, and then apply deep reinforcement learning algorithms to obtain effective decision strategies. Several experiments considering different confrontation scenarios are conducted to validate the feasibility and effectiveness of the DRL-based methods. Experimental results show that through learning the agents can make better decisions, and learning with MADDPG achieves superior performance than learning with other DRL-based models, which also explains the importance and necessity of mastering other agents' information.