See below for more details on how the GP model works. Alternative global optimization techniques like DIRECT or genetic algorithms are more flexible, but also typically require more evaluations than is feasible, especially in the presence of uncertainty.īayesian optimization starts by building a smooth surrogate model of the outcomes using Gaussian processes (GPs) based on the (possibly noisy) observations available from previous rounds of experimentation. However, grid search scales very poorly with the number of parameters (the dimensionality of the parameter space) and generally does not work well for more than a couple of continuous parameters. Parameter tuning is often done with simple strategies like grid search. The same strategy can be used to predict the expected gain from all future evaluations and decide on early termination, if the expected benefit is smaller than what is worthwhile for the problem at hand. BO is an adaptive approach where the observations from previous evaluations are used to decide what parameterizations to evaluate next. Tuning design parameters and rule-of-thumb heuristics for hardware design.īayesian optimization (BO) allows us to tune parameters in relatively few iterations by building a smooth model from an initial set of parameterizations (referred to as the "surrogate model") in order to predict the outcomes for as yet unexplored parameterizations.Finding optimal set of gait parameters for locomotive control in robotics, and.Hyperparameter optimization for machine learning,.Tuning Internet service parameters and selection of weights for recommender systems,.
These types of problems show up in a diversity of applications, such as When the number of parameters is not small or some of the parameters are continuous, using large factorial designs (e.g., “grid search”) or global optimization techniques for optimization require more evaluations than is practically feasible. In complex engineering problems we often come across parameters that have to be tuned using several time-consuming and noisy evaluations.
0 kommentar(er)
