Multiexperience-Assisted Efficient Multiagent Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Multiagent reinforcement learning (MARL) shows promise for cooperative policies in multiagent systems (MASs).
• Current MARL methods suffer from low sample efficiency, limiting real-world applications.
• Incorporating prior experience can accelerate MARL solution finding.

Purpose of the Study:

• To propose a novel multi-experience-assisted reinforcement learning (MEARL) method.
• To enhance the learning efficiency and performance of multiagent systems (MASs).
• To address the sample inefficiency drawback of traditional MARL.

Main Methods:

• Designed monotonicity-constrained reward shaping using expert experience for efficient multiagent guidance.
• Developed a reward distribution estimator utilizing transition experience to predict agent rewards.
• Estimated state value functions and accelerated convergence through reward prediction.

Main Results:

• MEARL demonstrated significant improvements in learning efficiency and overall performance in MASs.
• Evaluated performance on unmanned aerial vehicle combat (UAV-C) and StarCraft II Micromanagement (SCII-M) platforms.
• MEARL outperformed state-of-the-art methods in multiagent tasks.

Conclusions:

• The proposed MEARL method effectively enhances learning efficiency in MASs.
• MEARL offers a promising approach to overcome sample inefficiency in MARL.
• The method shows superior performance compared to existing state-of-the-art techniques.