Bayesian Optimization for Wide Landscapes (BOWL)

BOWL (Bayesian Optimization for Wide Landscapes) seeks robust optima—wide, flat basins that remain near-optimal under perturbations by combining Entropy-SGD with an MPD gradient oracle to guide black-box exploration.

Highlights

• Hybrid local-entropy objective steers search toward flat regions.
• MPD oracle ((v^* \propto -\Sigma^{-1}\mu)) picks descent directions from the GP posterior.
• Benchmarks (Hartmann-3D, Levy, Branin) show faster convergence and improved robustness against noise.

Limitations

• Not yet peer-reviewed—still in preprint form.
• High computational cost from the inner Entropy-SGD loop.
• Sensitive to tuning of Langevin-step sizes and GP update frequency.