Bayesian Optimization for Wide Landscapes (BOWL)

I spun up BOWL (Bayesian Optimization for Wide Landscapes - cool name ik!) as a researchy side project after reading Entropy-SGD and wondering how its “wide valleys” idea could be combined with Bayesian optimization in truly black-box settings. Since Bayesian “most probable descent” (MPD) gives you a way to approximate descent directions from a GP posterior (even without true gradients), you could theoretically plug that into an entropy-style inner loop and bias BayesOpt toward flat, robust basins instead of sharp, brittle optima.

I ended up building a hybrid algorithm where a GP + MPD-like gradient oracle feeds an Entropy-SGD–inspired Langevin inner loop, wrapped in a simple acquisition policy for global exploration. It seemed to work on classic toy problems/benchmarks in the BO space (Hartmann-3D, Levy, Branin) and does find good wide solutions, but it stayed mostly in low-dimensional, synthetic domains rather than becoming a fully practical optimizer.

For me, BOWL mainly showcases how I like to connect ideas across papers and prototype new hybrids, and it’s something I’d be excited to revisit and push further down the line.