A Multiphase Semistatic Training Method for Swarm Confrontation Using Multiagent Deep Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Swarm confrontation requires effective multi-agent training strategies.
• Traditional single-phase training can be inefficient for complex scenarios.
• Reinforcement learning offers a framework for agent training.

Purpose of the Study:

• To propose and evaluate a novel multiphase semistatic training method for swarm confrontation.
• To improve the training efficiency and learning outcomes for agents in competitive environments.
• To address the challenge of weak agents failing to learn in traditional training paradigms.

Main Methods:

• Development of a 3V3 tank fight game simulation using the Unity platform.
• Implementation of the Multi-Agent Proximal Policy Optimization with Curriculum Adaptation (MA-POCA) algorithm from the ML-Agent toolkit.
• Application of a multiphase learning strategy with incrementally increasing performance levels for strong agents.
• Incorporation of semistatic learning where strong agents pause learning against weaker opponents.

Main Results:

• The proposed multiphase semistatic training method significantly increases training efficiency compared to single-phase methods.
• Experimental results demonstrate improved learning outcomes for weaker agents.
• The method reduces the computational cost and time required for effective agent training.

Conclusions:

• The multiphase semistatic training method is an effective approach for swarm confrontation in multi-agent deep reinforcement learning.
• This strategy facilitates efficient knowledge transfer from strong to weak agents.
• The findings offer valuable insights for developing more capable and adaptable AI agents in competitive environments.