Graph-Based Deep Reinforcement Learning Approach for Alliance Formation Game based Robot Swarm Task Assignment
Yang Lv, Jinlong Lei, Peng Yi · 2024
This paper explores the robot swarm task allocation problem based on alliance formation game theory, which treats each robot as an autonomous agent capable of forming strategic alliances for task completion, with a focus on optimizing overall system revenue and mutual benefits. To resolve the problem, we introduce a unique graphbased Deep Reinforcement Learning (DRL) framework named AFGNet_DDQN. Firstly, we construct an Allocation Feature Graph (AFG) that intricately maps the complex interactive relationships and allocation features among robots and tasks, and develop the AFGNet architecture to efficiently extract features from the graph nodes. Then thorugh reconstructing a Markov Decision Process (MDP) within this graph, we implement an advanced version of Double Deep Q-Networks (DDQN) algorithm, adapted for our graph-based framework. This setup allows for the effective learning and optimization of task allocation strategies through localized interactions among the robots. Finally, our empirical results demonstrate the superiority of our framework over traditional methods, especially in terms of scalability and robustness.