The authors propose QADAPT, a framework for cooperative multi-agent reinforcement learning that addresses parameter cross-talk in electrostatically-defined quantum-dot arrays by using an online-learned factored representation of the action space to decouple agents.
- The method minimizes interference between agents through modular shared policies based on local measurements and rewards.
- It achieves zero-shot generalization to unseen quantum device sizes.
- The approach maintains an approximately constant number of convergence steps to reach target regimes regardless of scale.
This work provides a scalable route toward the rapid calibration of large-scale quantum processors.