Model Ensembling for Constrained Optimization

The paper presents two ensemble methods for decision-making under constrained optimization, focusing on model ensembling to improve performance in high-dimensional tasks. A white-box approach, requiring model access, ensures self-consistency and strong guarantees, while a black-box method addresses model independence and offers a swap regret-like guarantee. Both methods aim to enhance downstream optimization by leveraging multicalibration techniques. The study evaluates the algorithms in controlled experiments across various fields, comparing their convergence, utility, and efficiency. Key findings include convergence results, performance bounds, and empirical improvements over initial policies, with the white box method generally providing better outcomes but at a higher computational cost. The work also connects to related areas like decision calibration, prediction for actions, and extends prior research on ensemble methods for high-dimensional problems.