Textbook of AI

A multi-agent system (MAS) is a system of multiple agents, each pursuing its own objectives, interacting in a shared environment. Studied across AI, economics (game theory), distributed computing, robotics.

Key concepts:

• Cooperative MAS: agents share a common goal. Coordination problem dominant.
• Competitive MAS: agents have conflicting objectives. Game-theoretic analysis (Nash equilibrium, correlated equilibrium).
• Mixed-motive: cooperation on some dimensions, competition on others. Emergence of social norms.

Algorithms:

• Multi-agent reinforcement learning (MARL): extend RL to multiple learners. Independent Q-learning, joint-action learners, centralised training/decentralised execution (CTDE).
• Self-play: agents train against past versions of themselves. Used in AlphaGo Zero, AlphaZero, OpenAI Five (Dota 2), AlphaStar (StarCraft II).
• Population-based training: maintain a population of agents, train each against the population. Avoids cycles in self-play.
• Mean-field methods: approximate large populations by their statistical distribution.

In modern AI:

• LLM-based agents in collaboration: AutoGen, MetaGPT, CrewAI use multiple LLM-driven agents with different roles.
• AI safety multi-agent dynamics: collusion between AI systems, market manipulation by AI traders, recursive self-improvement.
• Constitutional AI: arguably a multi-agent dynamic, the model critiquing itself plays both critic and creator roles.

Multi-agent dynamics are an increasingly important area for AI safety research as more capable AI systems are deployed in the world together.