Learning Rational Strategies for Temporal Objectives in Multi-Agent Systems
The method seeks strategies that meet one agent’s probabilistic objective while keeping the joint strategy stable under a chosen equilibrium rule.
The researchers study systems with unknown dynamics, where each agent has a temporal objective and its payoff depends on the probability of satisfying it. They translate those objectives into reward automata for model-free reinforcement learning. The paper gives conditional convergence guarantees for tabular Nash-Q learning, finite-sample confidence bounds for objective satisfaction, and approximate best-response checks for strategic stability. The authors evaluate the approach on eight general-sum concurrent stochastic games.
ArXiv · AI/CL/LG's note
The researchers study systems with unknown dynamics, where each agent has a temporal objective and its payoff depends on the probability of satisfying it. They translate those objectives into reward automata for model-free reinforcement learning. The paper gives conditional convergence guarantees for tabular Nash-Q learning, finite-sample confidence bounds for objective satisfaction, and approximate best-response checks for strategic stability. The authors evaluate the approach on eight general-sum concurrent stochastic games.
ArXiv · AI/CL/LG's note
score 4