Resilient multi-agent RL: introducing DQ-RTS for distributed environments with data loss.

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Distributed systems face challenges with unreliable communication and dynamic agent populations.
• Existing Multi-Agent Reinforcement Learning (MARL) algorithms struggle in non-ideal, fluctuating environments.

Purpose of the Study:

• To propose DQ-RTS, a novel decentralized MARL algorithm.
• To enhance agent communication and adaptability in distributed settings.
• To evaluate DQ-RTS against existing methods in challenging scenarios.

Main Methods:

• Developed DQ-RTS with an optimized communication protocol.
• Conducted comparative analysis with Q-RTS (Q-learning for Real-Time Swarms).
• Performed extensive experiments on benchmark tasks with varying agent numbers and communication quality.

Main Results:

• DQ-RTS demonstrated superior convergence speed, with a speed-up factor of 1.6 to 2.7 under non-ideal communication.
• The algorithm maintained performance robustness despite fluctuating agent populations.
• Validated scalability and effectiveness across diverse benchmark tasks.

Conclusions:

• DQ-RTS offers a practical and resilient solution for MARL in dynamic distributed environments.
• The algorithm effectively addresses non-ideal communication and varying agent numbers.
• DQ-RTS shows significant improvements in convergence and robustness compared to Q-RTS.