Semicentralized Deep Deterministic Policy Gradient in Cooperative StarCraft Games.

Area of Science:

• Artificial Intelligence
• Reinforcement Learning
• Game Theory

Background:

• Cooperative multiagent systems face challenges in coordination and communication.
• Partially observable environments complicate decision-making for individual agents.
• Fully centralized methods can incur significant communication overhead.

Purpose of the Study:

• To propose a novel semicentralized deep deterministic policy gradient (SCDDPG) algorithm for cooperative multiagent games.
• To enhance agent interaction and cooperation in complex combat scenarios.
• To reduce communication burden compared to fully centralized approaches.

Main Methods:

• Designed a two-level actor-critic structure: local and global.
• Local actor-critic handles partially observable information for each agent type.
• Global actor-critic provides an overall combat view using limited centralized information.
• Developed tailored reward functions for local and global structures.

Main Results:

• The SCDDPG algorithm effectively enables agents to cooperate.
• Demonstrated improved performance in various StarCraft combat scenarios.
• Simulation results confirm successful teamwork and enemy defeat.

Conclusions:

• The proposed SCDDPG algorithm offers an efficient approach for cooperative multiagent games.
• The two-level structure balances local autonomy with global coordination.
• This method significantly reduces communication requirements while maintaining high performance.