Bayesian Optimization for Wide Landscapes (BOWL)

The idea behind is BOWL (Bayesian Optimization for Wide Landscapes) is to extend the Bayesian Optimization algorithm MPD to actively seek optima that exist in wide, flat basins (and hence robust) by combining Entropy-SGD with MPD’s gradient oracle to guide black-box exploration.

Highlights

• Introduces a hybrid local-entropy objective that steers soptimization toward flat, stable regions of the posterior surface.
• MPD oracle derived from the GP posterior suggests descent directions that reflect landscape uncertainty.
• Benchmarks (Hartmann-3D, Levy, Branin) show convergence and robustness against noise.

Limitations

• Not peer-reviewed; still in preprint form.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