Unilateral incentive alignment in two-agent stochastic games

Area of Science:

• Artificial Intelligence
• Game Theory
• Multiagent Systems

Background:

• Multiagent learning faces challenges due to mixed-motivation interactions and conflicting agent interests.
• Evolutionary game theory introduced zero-determinant (ZD) strategies, granting unilateral control over outcomes in repeated games.

Purpose of the Study:

• Generalize ZD strategies to stochastic games, enabling dynamic environment adaptation.
• Develop an algorithm for agents to discover strategies enforcing linear payoff relationships, including equal payoffs for fairness.

Main Methods:

• Proposed a novel algorithm for discovering generalized zero-determinant strategies.
• Focused on enforcing predetermined linear payoff relationships, particularly equal payoffs (fairness).
• Demonstrated strategy discovery through experience alone, without opponent coordination.

Main Results:

• Successfully generalized ZD strategies to stochastic games.
• An agent can learn to enforce linear payoff relationships, including fairness, via experience.
• The 'enforcer' strategy incentivizes optimal and equitable outcomes, preventing exploitation.

Conclusions:

• Generalized ZD strategies enhance multiagent learning in dynamic, mixed-motivation settings.
• Enforcer strategies can transform competitive interactions into cooperative ones, promoting fairness.
• This approach facilitates greater collaboration and equitable outcomes in multiagent systems.