BayGPGO 0.3.2

Bayesian Optimization with Gaussian Process as surrogate model

GPGO - Gaussian Process GO

My own implementation of a Bayesian Black box Optimization with Gaussian Process as a surrogate model. It is still in development but it was successfully used to achieve a bottom up optimization of the Dissipative Particle Dynamics force field for a complex system of polymers chains functionalized gold nanoparticles in a water solvent.

Hyperparameters

The Hyperparameters of the GP are optimized by the common technique of maximizing the Log Marginal Likelihood. In this repository this is achieved by using a search grid (although not in an efficient way) or by using the scipy optimizer module (L-BFGS-B, TNC, SLSCP). The analytical gradient is implemented for the Radial Basis Function kernel and it is possible to use the derivate of the Log Marginal Likelihood to optimize the hyperparameters.

Acquisition function

As it is there are two different acquisition function implemented right now:

-Expected Improvement (EI)

-UCB (Upper Confidence Bound)

Maximizing the Acquisition function

In this little package right now there are 3 ways to run an optimization task with Gaussian Processes:

-NAIVE : AkA sampling the acquisition function with a grid of some kind or a quasi random methods as LHS (require smt package)

-BFGS : optimize the Acquisition function by using the L-BFGS-B optimizer

-DIRECT : optimize the Acquisition function by using the DIRECT optimizer (require DIRECT python package)

-PSO : optimize the Acquisition function by a Particle Swarm Optimization genetic algorithm

Made for experiments

Easy to use it with a shell procedure! Load the data and just .suggest_location() to get the next points of your experiment!

Multi Objective

Right now the package contains an implementation of the NSGAII genetic solver that allows to solve multi objective problems. It has also an early version of a Multi Objective Bayesian optimization that uses the NSGAII and optimize the Acquisition function EI or the mean function of the GP. It will follow a more precise implementation of the Hypervolume improvement.

TODO

-Load data routines

-Good code practice maybe

-Easy routines for LAMMPS (at least DPD and BD)