Mechanisms for Energy Efficient AI Training

This poster addresses the rising energy footprint of distributed and collaborative artificial intelligence (AI) systems. It introduces a game-theoretic framework designed to align the motivations of self-interested, local training participants with collective environmental goals. In a federated learning environment where users train models locally, a server aims to manage energy consumption against shared green energy reserves while balancing individual user efforts, performance, and costs. The authors propose and evaluate two coordination mechanisms: a baseline heuristic phase based on independent client decision-making, and an advanced correlation device where an external source provides optimisation signals to maximise green energy use. Evaluated on synthetic data, the correlation device demonstrated superior performance, consistently using green energy more efficiently than standard Nash equilibrium outcomes and yielding a smoother, more adaptive consumption profile without degrading the AI model's learning outcomes.