Bayesian optimization for evalution of posterior distributions
Výpočet aposteriorní distribuce Bayesovskou optimalizací
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učení technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date
Abstract
V mnoha odvětvích vědy, inženýrství a ekonomie se komplexní systémy modelují pomocí počítačových simulací, které poskytují neocenitelný vhled do jejich chování. Simulace se používají k inferenci neznámých vlastností těchto systémů. Inference se obvykle provádí zkoušením různých hodnot neznámých parametrů a ověřením jejich shody s realitou potom, co je simulace dokončena. Často jsou ale simulace výpočetně náročné, což proces pokus-omyl dělá nevhodným. Tento problém motivuje vznik metod zvaných simulation-based inference (SBI), které usilují o automatizaci tohoto procesu, poskytnutí přesné inference neznámých parametrů společně s jejich nejistotou a snížení počtu potřebných simulací. V této diplomové práci formálně definujeme SBI problém a představíme nové rozšíření tohoto problému nazvané simulation-based feature selection (SBFS). SBFS problém se zabývá výběrem nejlepších rysů, podle kterých provádět inferenci. Představíme metodu BOLFI, jako řešení SBI založené na Bayesovké optimalizaci (BO), která modeluje neshodu simulace s realitou pomocí Gausovského procesu a aktivně volí simulace tak, aby se maximalizovala jejich užitečnost. Rozšíříme metodu BOLFI pro SBFS a ukážeme její přesnost a efektivitu na analytickém vzorovém příkladu a na reálném problému simulace plasmatu v tokamaku.
In many fields of science, engineering, and economy, complex systems are modeled with computer simulations, which provide invaluable insight into their behavior. The simulations are used to infer unknown properties of the real system. This inference is commonly performed by guessing the values of the parameters of interest, and verifying their agreement with reality once the simulation concludes. However, it is often the case that the simulations are computationally demanding, making this process of trial-and-error bothersome. This motivates the field of simulation-based inference (SBI), which aims to automate this process, provide accurate inference of the parameters together with their uncertainty, and limit the number of required simulations. In this thesis, we formally define the SBI problem, and we introduce a new extension of the problem, dubbed the simulator-based feature selection (SBFS). The SBFS problem concerns the selection of the best features to base the inference on. We present the BOLFI method, a SBI approach based on the Bayesian optimization (BO), which models the discrepancies of the simulation and reality via Gaussian processes, and actively selects simulations to maximize their utility. We extend the BOLFI method for SBFS, and showcase its accuracy and efficiency on an analytical toy problem as well as a real-world problem of simulation of plasma inside a tokamak.
In many fields of science, engineering, and economy, complex systems are modeled with computer simulations, which provide invaluable insight into their behavior. The simulations are used to infer unknown properties of the real system. This inference is commonly performed by guessing the values of the parameters of interest, and verifying their agreement with reality once the simulation concludes. However, it is often the case that the simulations are computationally demanding, making this process of trial-and-error bothersome. This motivates the field of simulation-based inference (SBI), which aims to automate this process, provide accurate inference of the parameters together with their uncertainty, and limit the number of required simulations. In this thesis, we formally define the SBI problem, and we introduce a new extension of the problem, dubbed the simulator-based feature selection (SBFS). The SBFS problem concerns the selection of the best features to base the inference on. We present the BOLFI method, a SBI approach based on the Bayesian optimization (BO), which models the discrepancies of the simulation and reality via Gaussian processes, and actively selects simulations to maximize their utility. We extend the BOLFI method for SBFS, and showcase its accuracy and efficiency on an analytical toy problem as well as a real-world problem of simulation of plasma inside a tokamak.