Consensus of Multi-agent Reinforcement Learning Systems: The Effect of Immediate Rewards

This paper studies the consensus problem of a leaderless, homogeneous, multi-agent reinforcement learning (MARL) system using actor-critic algorithms with and without malicious agents. The goal each agent is to reach position maximum cumulative reward. Although reward function converges in both scenarios, absence agent, higher than present. We consider here various immediate functions. First, we study based on Manhattan distance. In addition proposing three different functions Euclidean, $n$-norm, Chebyshev distances, have rigorously shown which method has better performance for entire team Finally, present combination that yields By increasing agents’ combined function, demonstrated presence comparable agent. claims been proven theoretically, simulation confirms theoretical findings.