New AI framework achieves 99% satellite scheduling efficiency via iterative pricing

A team from CMU and NASA JPL (Itai Zilberstein, Pranav Rajbhandari, Steve Chien, Tuomas Sandholm) has developed a novel distributed constraint optimization (DCOP) framework for large-scale satellite scheduling. The challenge: coordinating dozens of satellites to fulfill observation requests under limited communication and tight constraints. Traditional DCOP solvers fail at this scale. The researchers decomposed the problem into a high-level meta-DCOP for task allocation and independent local scheduling subproblems. They introduced an iterative pricing method that dynamically adjusts the meta-level utilities based on feedback from local optimizers, coupling the two levels without centralizing.

To solve the meta-DCOP efficiently, they adapted modern online learning algorithms originally designed for potential games—showing these methods perform competitively with incomplete DCOP solvers. On real-world decentralized satellite scheduling instances, the combined approach fulfilled over 99% of observation requests, compared to only 87% for state-of-the-art baselines. The paper (arXiv:2607.25835) demonstrates that this framework scales to networks with hundreds of satellites and thousands of requests, making it practical for next-generation Earth observation constellations. The authors provide both theoretical grounding and empirical validation, marking a significant step in applying game-theoretic online learning to constrained optimization at scale.