Publication

Abstract : The growing complexity of decision-making in dynamic, decentralized settings requires more efficient, scalable multi-agent systems. Traditional methods of a centralized decision-making approach have proven to be rather slow to adapt in real-time and to scale effectively toward highly performing adversarial environments. This research aims to develop a decentralized multi-agent system that provides an efficient solution to decision-making by combining policy learning through deep reinforcement learning to allow agents to make optimized decisions and market-based task allocation mechanisms. To realize this, a multi-agent reinforcement learning (MARL) framework is set up where each agent independently learns the optimal policy based on local observations. The proposed methods allowed for significant improvements in efficiency in all aspects of the system, including up to 33.7% improvement in task completion time and 35.5% reduction in decision-making latency over baseline systems. These results demonstrate that the proposed system is robust and scalable to a variety of cooperative and competitive settings. This research provides useful information for developing more efficient, decentralized systems and techniques for real-world tasks such as autonomous systems, robotics, and distributed networks. Future research should explore further scalability and integration of ethical and regulatory considerations for real-world applications.