Online Sensor Grouping via Multi-Agent Learning Automata: An Ising Model Perspective
Abstract:
We introduce a novel framework for sensor grouping using Learning Automata, drawing inspiration from the Ising model in statistical physics. By representing binary sensor actions as atomic spins, we propose to model the problem as a common Ising-type separation game with interpretable utility functions that promote alignment or anti-alignment based on sensor reliability.
In particular, our study presents four reward realizations of this game scenario: ferromagnetic, anti-ferromagnetic, local ferromagnetic, and local anti-ferromagnetic. We rigorously establish that each realization constitutes a potential game aligned with the same separation potential, ensuring that individual sensor incentives contribute to the same overarching separation objective.
We also establish that, under standard small-step Learning Automata updates,the induced learning dynamics converge (in the stochastic-approximation sense) to pure-strategy Nash equilibria of the corresponding potential game, which are local maximizers of the separation potential. Under the standard two-type reliability model (with pi not equal to one-half), these equilibria coincide with the global maximizers and provide perfect separation of fair and unfair sensors (up to a global label flip).
Our proposed methods do not require ground truth or local stationarity assumptions about the processes measured by the sensors. By construction, our algorithm uses only current observations and requires minimal state memory in each sensor, making it suitable for near-real-time applications.
This work also paves the way for future research in multi-agent coordination leveraging physics principles under noisy feedback conditions, while providing an interpretable framework.
Citation:
A. Yazidi, M. Pinto-Orellana, and Y. Djenouri. 2026. Online Sensor Grouping via Multi-Agent Learning Automata: An Ising Model Perspective. To appear in Proc. of the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026), Paphos, Cyprus, May 25–29, 2026. IFAAMAS.