PSOPF-MATD3: A multi-agent collaborative radioactive source search strategy

This paper addresses the problem of multi-agent collaborative radioactive source localization using multi-agent deep reinforcement learning (MADRL). In this problem, agents need to learn collaborative search and collision-avoidance behaviors. Therefore, we propose a source search strategy, abbreviated as PSOPF-MATD3, that combines the particle swarm optimized particle filter (PSOPF) algorithm with the multi-agent twin delayed deep deterministic policy gradient (MATD3) algorithm. Specifically, the PSOPF algorithm is used to estimate the state of the source, and the key features of the source items are extracted using the Gaussian mixture model as inputs to the neural network. The MATD3 algorithm is used to find the optimal source search strategy based on the estimated status of the source terms. Experimental results show that the proposed PSOPF-MATD3 algorithm outperforms two multi-agent reinforcement learning algorithms (MADDPG and MASAC) in general and difficult scenarios. The proposed algorithm demonstrates a higher average group reward during the evaluation phase, completes the multi-agent collaborative target detection task in a shorter time, and exhibits superior convergence performance.