← Back to projects

Ongoing Research

Statistical Mechanics of Cooperative Games

Exploring cooperative game valuation through entropy-based modeling, variational inference, and statistical mechanics.

  • Cooperative Game Theory
  • Statistical Mechanics
  • Variational Inference

Ongoing research · Unpublished work

Overview

This ongoing project explores cooperative game valuation through the lens of statistical mechanics. It brings together entropy-based modeling, variational inference, and the study of collective behavior to investigate the foundations of valuation methods relevant to machine learning.

Background

Cooperative game theory provides a framework for assigning value to participants whose contributions depend on collaboration. In machine learning, this perspective has been applied to valuing training data, as illustrated by Data Shapley (Ghorbani and Zou, ICML 2019).

An energy-based formulation offers a connection to statistical physics: coalitions are treated as configurations of a probabilistic system, and valuation is studied through variational inference. Bian et al. (ICLR 2022) develop this framework, relating established game-theoretic valuation criteria to mean-field inference and introducing the Variational Index.

Research Questions

How does the structure of cooperation influence the behavior of an energy-based valuation system? How should collective effects, stability, and sensitivity to modeling assumptions be understood? When does a simplified approximation faithfully represent the underlying coalition model?

These questions motivate a statistical-mechanical investigation of valuation, with particular attention to the distinction between properties of the original probabilistic model and those introduced by approximation.

Approach

The project combines analytical study of representative cooperative games with numerical exploration. Entropy, free-energy landscapes, symmetry, and mean-field methods provide complementary tools for examining how individual contributions relate to collective behavior.

The broader aim is to develop a clearer physical understanding of cooperative valuation and its potential relevance to machine-learning applications.

Ongoing research · Unpublished work. This page presents the research motivation and general approach; unpublished theoretical results, proofs, and numerical findings are not included.