MATS

MATS implements a Bayesian Thompson Sampling algorithm to coordinate loosely-coupled cooperative agents in sparse multi-agent systems by balancing exploration and exploitation to learn coordinated policies.

Key Features:

• Bayesian Framework: Employs a Bayesian approach (Thompson Sampling) to balance exploration and exploitation for learning optimal strategies under uncertainty.
• Sparse Coordination Graphs: Models systems with sparse neighborhood structures where each agent's actions primarily affect its immediate neighbors to reduce coordination complexity.
• Regret Bound: Provides a theoretical regret bound that is sublinear in time and a low-order polynomial in the maximum number of actions available to any single agent.
• Empirical Performance: Demonstrates performance superior to MAUCE across benchmarks including synthetic tasks and scenarios with Poisson distributions.

Scientific Applications:

• Wind Farm Control: Learns coordinated turbine alignment policies that exploit upstream–downstream sparse interactions to maximize power production.
• Optimization of Power Production: Applied to realistic wind farm control tasks to achieve significant improvements over existing coordination methods.

Methodology:

Implements Multi-Agent Thompson Sampling within a Bayesian exploration–exploitation framework, models sparse interactions via coordination graphs, derives sublinear-in-time and low-order-polynomial regret bounds, and evaluates empirically against MAUCE on synthetic and Poisson-distribution scenarios.