An Improved Approach towards Multi-Agent Pursuit-Evasion Game Decision-Making Using Deep Reinforcement Learning

Area of Science:

• Robotics
• Artificial Intelligence
• Control Theory

Background:

• Pursuit-evasion games are a classic challenge in multi-agent systems (MASs).
• Traditional methods require complex state variables for modeling.
• Environment sensing and decision-making are critical for effective gameplay.

Purpose of the Study:

• To develop an online decision technique using deep reinforcement learning (DRL) for pursuit-evasion games.
• To implement a cooperative decision-making framework overcoming traditional modeling limitations.
• To enhance the robustness of agents against real-world uncertainties and model errors.

Main Methods:

• A control-oriented framework based on the multi-agent deep deterministic policy gradient (MADDPG) algorithm was developed.
• A novel adversarial attack trick and adversarial learning MADDPG (A2-MADDPG) algorithm were proposed.
• Adversarial learning was incorporated to model real-world uncertainties and optimize robust training.

Main Results:

• The proposed A2-MADDPG algorithm demonstrated superior performance for both pursuers and evaders.
• Agents effectively learned corresponding confrontational strategies during training.
• The approach successfully modeled real-world uncertainties, optimizing robust training.

Conclusions:

• The developed A2-MADDPG approach provides an effective solution for pursuit-evasion games in MASs.
• Adversarial learning enhances agent robustness and decision-making in dynamic environments.
• The framework enables cooperative decision-making without complex state variable modeling.