DOCSLIB.ORG

Computer Experimental Design for Gaussian Process Surrogates

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Computer Experimental Design for Gaussian Process Surrogates Computer Experimental Design for Gaussian Process Surrogates Boya Zhang Dissertation submitted to the Faculty of the Virginia Polytechnic Institute and State University in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Statistics Robert B. Gramacy, Chair Xinwei Deng David Higdon Leanna House August 13, 2020 Blacksburg, Virginia Keywords: Computer experiment, experimental design, sequential design, Gaussian process surrogates, input-dependent noise Copyright 2020, Boya Zhang Computer Experimental Design for Gaussian Process Surrogates Boya Zhang (ABSTRACT) With a rapid development of computing power, computer experiments have gained popu- larity in various scientific fields, like cosmology, ecology and engineering. However, some computer experiments for complex processes are still computationally demanding. A surro- gate model or emulator, is often employed as a fast substitute for the simulator. Meanwhile, a common challenge in computer experiments and related fields is to efficiently explore the input space using a small number of samples, i.e., the experimental design problem. This dissertation focuses on the design problem under Gaussian process surrogates. The first work demonstrates empirically that space-filling designs disappoint when the model hyperparame- terization is unknown, and must be estimated from data observed at the chosen design sites. A purely random design is shown to be superior to higher-powered alternatives in many cases. Thereafter, a new family of distance-based designs are proposed and their superior performance is illustrated in both static (one-shot design) and sequential settings. The sec- ond contribution is motivated by an agent-based model(ABM) of delta smelt conservation. The ABM is developed to assist in a study of delta smelt life cycles and to understand sen- sitivities to myriad natural variables and human interventions. However, the input space is high-dimensional, running the simulator is time-consuming, and its outputs change nonlin- early in both mean and variance. A batch sequential design scheme is proposed, generalizing one-at-a-time variance-based active learning, as a means of keeping multi-core cluster nodes fully engaged with expensive runs. The acquisition strategy is carefully engineered to favor selection of replicates which boost statistical and computational efficiencies. Design per- formance is illustrated on a range of toy examples before embarking on a smelt simulation campaign and downstream high-fidelity input sensitivity analysis. Computer Experimental Design for Gaussian Process Surrogates Boya Zhang (GENERAL AUDIENCE ABSTRACT) With a rapid development of computing power, computer experiments have gained popu- larity in various scientific fields, like cosmology, ecology and engineering. However, some computer experiments for complex processes are still computationally demanding. Thus, a statistical model built upon input-output observations, i.e., a so-called surrogate model or emulator, is needed as a fast substitute for the simulator. Design of experiments, i.e., how to select samples from the input space under budget constraints, is also worth studying. This dissertation focuses on the design problem under Gaussian process (GP) surrogates. The first work demonstrates empirically that commonly-used space-filling designs disappoint when the model hyperparameterization is unknown, and must be estimated from data ob- served at the chosen design sites. Thereafter, a new family of distance-based designs are proposed and their superior performance is illustrated in both static (design points are al- located at one shot) and sequential settings (data are sampled sequentially). The second contribution is motivated by a stochastic computer simulator of delta smelt conservation. This simulator is developed to assist in a study of delta smelt life cycles and to understand sensitivities to myriad natural variables and human interventions. However, the input space is high-dimensional, running the simulator is time-consuming, and its outputs change nonlin- early in both mean and variance. An innovative batch sequential design method is proposed, generalizing one-at-a-time sequential design to one-batch-at-a-time scheme with the goal of parallel computing. The criterion for subsequent data acquisition is carefully engineered to favor selection of replicates which boost statistical and computational efficiencies. The design performance is illustrated on a range of toy examples before embarking on a smelt simulation campaign and downstream sensitivity analysis of simulator inputs. Dedication To my dearest family. iv Acknowledgments This dissertation would not have been possible without the help and support of my advisor, committee, colleagues, friends and family. First, I would like to express my deepest grati- tude to my advisor Bobby Gramacy. With his guidance, I have opportunities to learn about many interesting and challenging research areas, which motivates me to explore more in the future. In the past three years, he gave me persistent help and encouragement in overcoming difficulties. I feel so grateful to have the chance of working with him. I want to thank my committee members Dr. Xinwei Deng, Dr. Dave Higdon, and Dr. Leanna House. Their valuable advice and thought-provoking questions have broadened my view and pushed me to think deeper. I would also like to thank the statistics department of Virginia Tech for providing various courses on cutting-edge topics and plenty of opportuni- ties in teaching and collaborating. And my biggest thanks to all my friends, especially Ruijin Lu, who saved me from home- lessness during my dessertation writing period. Last but not least, I would like to thank my family, in particular, my parents, Jinling Ma and Yongchun Zhang, my grandparents, Qinglan Wu, Meiying Ni and Shimin Zhang, for their unconditional love. Their support gives me the strength to face any difficulties. v Contents List of Figures x List of Tables xiii 1 Introduction 1 1.1 Background .................................... 1 1.2 Motivation example: Delta Smelt ........................ 2 1.3 Overview of this dissertation ........................... 3 1.3.1 Distance-distributed design for GP surrogates ............. 3 1.3.2 IMSPE batch sequential design ..................... 3 1.3.3 Delta Smelt ................................ 4 2 Review of literature 5 2.1 Surrogate modeling ................................ 5 2.1.1 Gaussian Process surrogates ....................... 6 2.1.2 GP kernels ................................ 7 2.1.3 Surrogates for stochastic computer simulators ............. 9 2.1.4 GP with replication ........................... 10 2.1.5 Heteroskedastic Gaussian process .................... 11 vi 2.2 Computer experimental design .......................... 12 2.2.1 Geometric designs ............................ 13 2.2.2 Model-based design ............................ 15 2.2.3 Sequential design ............................. 16 2.2.4 Bayesian Optimization .......................... 18 2.2.5 Batch sequential design ......................... 19 3 Distance-distributed Design for Gaussian Process Surrogates 21 3.1 Setup and related work .............................. 24 3.1.1 Gaussian Process surrogates ....................... 25 3.1.2 Thinking about designs for GPs ..................... 25 3.2 Better than random ............................... 28 3.2.1 Uniform to beta designs ......................... 31 3.2.2 Optimization of shape parameters of betadist design ......... 34 3.3 Hybrid betadist and LHS ............................ 39 3.4 Application to sequential design ......................... 41 3.4.1 Active Learning MacKay ......................... 42 3.4.2 Expected improvement for optimization ................ 44 4 IMSPE batch-sequential design 48 4.1 Batch sequential design .............................. 50 vii 4.1.1 A criterion for minimizing variance ................... 50 4.1.2 Batch IMSPE gradient .......................... 53 4.1.3 Implementation details and illustration ................. 57 4.2 Hunting for replicates .............................. 59 4.2.1 Backtracking via merge ......................... 59 4.2.2 Selecting among backtracked batches .................. 61 4.3 Benchmarking examples ............................. 62 4.3.1 1d toy example .............................. 64 4.3.2 2d toy example .............................. 64 4.3.3 Ocean oxygen ............................... 68 4.3.4 Assemble-to-order ............................ 70 5 Delta smelt 72 5.1 Agent-based model ................................ 74 5.2 Pilot study .................................... 77 5.3 Big experiment .................................. 79 5.3.1 Setup and acquisitions .......................... 80 5.3.2 Downstream analysis ........................... 83 6 Conclusion 88 6.1 Distance-distributed design for GP surrogates ................. 88 viii 6.2 IMSPE Batch-sequential design ......................... 90 6.3 Delta Smelt simulator .............................. 93 Bibliography 94 ix List of Figures 2.1 GP posterior predictive distribution in terms of means, 2.5% and 97.5% quan- tiles. ....................................... 8 2.2 Predictions and associated 95% uncertainty intervals based on GP with nugget parameter. ..................................... 10 3.1 logMSEs from design experiment and de-trending surface. .......... 28 3.2 Standardized logMSE boxplots to thirty gridded θ(t) values for seven compara- tors using n = 2d+1
Load more

Recommended publications
  • An Uncertainty-Quantification Framework for Assessing Accuracy
    An Uncertainty-Quantification Framework for Assessing Accuracy, Sensitivity, and Robustness in Computational Fluid Dynamics S. Rezaeiravesha,b,∗, R. Vinuesaa,b,∗∗, P. Schlattera,b,∗∗ aSimEx/FLOW, Engineering Mechanics, KTH Royal Institute of Technology, SE-100 44 Stockholm, Sweden bSwedish e-Science Research Centre (SeRC), Stockholm, Sweden Abstract A framework is developed based on different uncertainty quantification (UQ) techniques in order to assess validation and verification (V&V) metrics in computational physics problems, in general, and computational fluid dynamics (CFD), in particular. The metrics include accuracy, sensitivity and robustness of the simula- tor's outputs with respect to uncertain inputs and computational parameters. These parameters are divided into two groups: based on the variation of the first group, a computer experiment is designed, the data of which may become uncertain due to the parameters of the second group. To construct a surrogate model based on uncertain data, Gaussian process regression (GPR) with observation-dependent (heteroscedastic) noise structure is used. To estimate the propagated uncertainties in the simulator's outputs from first and also the combination of first and second groups of parameters, standard and probabilistic polynomial chaos expansions (PCE) are employed, respectively. Global sensitivity analysis based on Sobol decomposition is performed in connection with the computer experiment to rank the parameters based on their influence on the simulator's output. To illustrate its capabilities, the framework is applied to the scale-resolving simu- lations of turbulent channel flow using the open-source CFD solver Nek5000. Due to the high-order nature of Nek5000 a thorough assessment of the results' accuracy and reliability is crucial, as the code is aimed at high-fidelity simulations.
    [Show full text]
  • Designing Combined Physical and Computer Experiments to Maximize Prediction Accuracy
    Computational Statistics and Data Analysis 113 (2017) 346–362 Contents lists available at ScienceDirect Computational Statistics and Data Analysis journal homepage: www.elsevier.com/locate/csda Designing combined physical and computer experiments to maximize prediction accuracy Erin R. Leatherman a,∗, Angela M. Dean b, Thomas J. Santner b a West Virginia University, United States b The Ohio State University, United States article info a b s t r a c t Article history: Combined designs for experiments involving a physical system and a simulator of the Received 1 March 2016 physical system are evaluated in terms of their accuracy of predicting the mean of Received in revised form 18 July 2016 the physical system. Comparisons are made among designs that are (1) locally optimal Accepted 19 July 2016 under the minimum integrated mean squared prediction error criterion for the combined Available online 25 July 2016 physical system and simulator experiments, (2) locally optimal for the physical or simulator experiments, with a fixed design for the component not being optimized, (3) maximin Keywords: augmented nested Latin hypercube, and (4) I-optimal for the physical system experiment Augmented nested LHD Calibration and maximin Latin hypercube for the simulator experiment. Computational methods are Gaussian process Kriging interpolator proposed for constructing the designs of interest. For a large test bed of examples, the IMSPE empirical mean squared prediction errors are compared at a grid of inputs for each test Latin hypercube design surface using a statistically calibrated Bayesian predictor based on the data from each Simulator design. The prediction errors are also studied for a test bed that varies only the calibration parameter of the test surface.
    [Show full text]
  • Computational Physics: Simulation of Classical and Quantum Systems
    Computational Physics Philipp O.J. Scherer Computational Physics Simulation of Classical and Quantum Systems 123 Prof. Dr. Philipp O.J. Scherer TU München Physikdepartment T38 85748 München Germany [email protected] Additional materials to this book can be downloaded from http://extras.springer.com ISBN 978-3-642-13989-5 e-ISBN 978-3-642-13990-1 DOI 10.1007/978-3-642-13990-1 Springer Heidelberg Dordrecht London New York Library of Congress Control Number: 2010937781 c Springer-Verlag Berlin Heidelberg 2010 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover design: eStudio Calamar S.L., Heidelberg Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com) for Christine Preface Computers have become an integral part of modern physics. They help to acquire, store, and process enormous amounts of experimental data.
    [Show full text]
  • Calibrating a Large Computer Experiment Simulating Radiative
    The Annals of Applied Statistics 2015, Vol. 9, No. 3, 1141–1168 DOI: 10.1214/15-AOAS850 c Institute of Mathematical Statistics, 2015 CALIBRATING A LARGE COMPUTER EXPERIMENT SIMULATING RADIATIVE SHOCK HYDRODYNAMICS By Robert B. Gramacy∗, Derek Bingham†, James Paul Holloway‡, Michael J. Grosskopf‡, Carolyn C. Kuranz‡, Erica Rutter‡, Matt Trantham‡ and R. Paul Drake‡ University of Chicago∗, Simon Fraser University† and University of Michigan‡ We consider adapting a canonical computer model calibration apparatus, involving coupled Gaussian process (GP) emulators, to a computer experiment simulating radiative shock hydrodynamics that is orders of magnitude larger than what can typically be ac- commodated. The conventional approach calls for thousands of large matrix inverses to evaluate the likelihood in an MCMC scheme. Our approach replaces that costly ideal with a thrifty take on essential ingredients, synergizing three modern ideas in emulation, calibration and optimization: local approximate GP regression, modularization, and mesh adaptive direct search. The new methodology is motivated both by necessity—considering our particular application—and by re- cent trends in the supercomputer simulation literature. A synthetic data application allows us to explore the merits of several variations in a controlled environment and, together with results on our mo- tivating real-data experiment, lead to noteworthy insights into the dynamics of radiative shocks as well as the limitations of the calibra- tion enterprise generally. 1. Introduction. Rapid increases in computational power have made com- puter models (or simulators) commonplace as a way to explore complex physical systems, particularly as an alternative to expensive field work or physical arXiv:1410.3293v4 [stat.AP] 5 Nov 2015 experimentation.
    [Show full text]
  • Sequential Design of Computer Experimentsfor The
    Sequential design of computer experiments for the estimation of a probability of failure Julien Bect⋆ · David Ginsbourger · Ling Li · Victor Picheny · Emmanuel Vazquez⋆ Received: date / Accepted: date Abstract This paper deals with the problem of estimating the volume of the excursion set of a function f : Rd R above a given threshold, under a probability measure on Rd that is assumed to be known. In → the industrial world, this corresponds to the problem of estimating a probability of failure of a system. When only an expensive-to-simulate model of the system is available, the budget for simulations is usually severely limited and therefore classical Monte Carlo methods ought to be avoided. One of the main contributions of this article is to derive SUR (stepwise uncertainty reduction) strategies from a Bayesian-theoretic formulation of the problem of estimating a probability of failure. These sequential strategies use a Gaussian process model of f and aim at performing evaluations of f as efficiently as possible to infer the value of the probability of failure. We compare these strategies to other strategies also based on a Gaussian process model for estimating a probability of failure. Keywords Computer experiments Sequential design Gaussian processes Probability of failure · · · · Stepwise uncertainty reduction 1 Introduction The design of a system or a technological product has to take into account the fact that some design parameters are subject to unknown variations that may affect the reliability of the system. In particular, it is important to estimate the probability of the system to work under abnormal or dangerous operating conditions due to random dispersions of its characteristic parameters.
    [Show full text]
  • Design of Computer Experiments: Space Filling and Beyond Luc Pronzato, Werner Müller
    Design of computer experiments: space filling and beyond Luc Pronzato, Werner Müller To cite this version: Luc Pronzato, Werner Müller. Design of computer experiments: space filling and beyond. Statistics and Computing, Springer Verlag (Germany), 2012, 22 (3), pp.681-701. 10.1007/s11222-011-9242-3. hal-00685876 HAL Id: hal-00685876 https://hal.archives-ouvertes.fr/hal-00685876 Submitted on 6 Apr 2012 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. Statistics and Computing manuscript No. (will be inserted by the editor) Design of computer experiments: space ¯lling and beyond Luc Pronzato ¢ Werner G. MÄuller January 28, 2011 Abstract When setting up a computer experiment, it 1 Introduction has become a standard practice to select the inputs spread out uniformly across the available space. These Computer simulation experiments (see, e.g., Santner so-called space-¯lling designs are now ubiquitous in cor- et al (2003); Fang et al (2005); Kleijnen (2009)) have responding publications and conferences. The statisti- now become a popular substitute for real experiments cal folklore is that such designs have superior properties when the latter are infeasible or too costly. In these when it comes to prediction and estimation of emula- experiments, a deterministic computer code, the sim- tor functions.
    [Show full text]
  • Computer Experiments
    S. Ghosh and C. R. Rao, eds., Handbook of Statistics, Vol. 13 N © 1996 Elsevier Science B.V. All rights reserved. Computer Experiments J. R. Koehler and A. B. Owen 1. Introduction Deterministic computer simulations of physical phenomena are becoming widely used in science and engineering. Computers are used to describe the flow of air over an airplane wing, combustion of gases in a flame, behavior of a metal structure under stress, safety of a nuclear reactor, and so on. Some of the most widely used computer models, and the ones that lead us to work in this area, arise in the design of the semiconductors used in the computers themselves. A process simulator starts with a data structure representing an unprocessed piece of silicon and simulates the steps such as oxidation, etching and ion injection that produce a semiconductor device such as a transistor. A device simulator takes a description of such a device and simulates the flow of current through it under varying conditions to determine properties of the device such as its switching speed and the critical voltage at which it switches. A circuit simulator takes a list of devices and the way that they are arranged together and computes properties of the circuit as a whole. In each of these computer simulations, the user must specify the values of some governing variables. For example in process simulation the user might have to specify the duration and temperature of the oxidation step, and doses and energies for each of the ion implantation steps. These are continuously valued variables.
    [Show full text]
  • Sequential Experimental Designs for Stochastic Kriging
    Proceedings of the 2014 Winter Simulation Conference A. Tolk, S. Y. Diallo, I. O. Ryzhov, L. Yilmaz, S. Buckley, and J. A. Miller, eds. SEQUENTIAL EXPERIMENTAL DESIGNS FOR STOCHASTIC KRIGING Xi Chen Qiang Zhou Industrial and Systems Engineering Systems Engineering and Engineering Management Virginia Tech City University of Hong Kong Blacksburg, VA 24601, USA Kowloon Tong, HK ABSTRACT Recently the stochastic kriging (SK) methodology proposed by Ankenman et al. (2010) has emerged as an effective metamodeling tool for approximating a mean response surface implied by a stochastic simulation. Although fruitful results have been achieved through bridging applications and theoretical investigations of SK, there lacks a unified account of efficient simulation experimental design strategies for applying SK metamodeling techniques. In this paper, we propose a sequential experimental design framework for applying SK to predicting performance measures of complex stochastic systems. This framework is flexible; i.e., it can incorporate a variety of design criteria. We propose several novel design criteria under the proposed framework, and compare the performance with that of classic non-sequential designs. The evaluation uses illustrative test functions and the well-known M/M/1 and the (s;S) inventory system simulation models. 1 INTRODUCTION To build a high-quality metamodel with a given computational budget to expend, a carefully designed simu- lation experiment is critical. The literature on experimental designs for deterministic computer experiments abounds and various design schemes have been proposed, for instance, Latin hypercube designs (LHDs) (McKay et al. 1979), orthogonal array based LHDs (Tang 1993), uniform designs (Fang et al. 2000), and maximum entropy designs (Shewry and Wynn 1987), to name a few.
    [Show full text]
  • Design and Analysis of Computer Experiments for Screening Input Variables Dissertation Presented in Partial Fulfillment of the R
    Design and Analysis of Computer Experiments for Screening Input Variables Dissertation Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University By Hyejung Moon, M.S. Graduate Program in Statistics The Ohio State University 2010 Dissertation Committee: Thomas J. Santner, Co-Adviser Angela M. Dean, Co-Adviser William I. Notz ⃝c Copyright by Hyejung Moon 2010 ABSTRACT A computer model is a computer code that implements a mathematical model of a physical process. A computer code is often complicated and can involve a large number of inputs, so it may take hours or days to produce a single response. Screening to determine the most active inputs is critical for reducing the number of future code runs required to understand the detailed input-output relationship, since the computer model is typically complex and the exact functional form of the input- output relationship is unknown. This dissertation proposes a new screening method that identifies active inputs in a computer experiment setting. It describes a Bayesian computation of sensitivity indices as screening measures. It provides algorithms for generating desirable designs for successful screening. The proposed screening method is called GSinCE (Group Screening in Computer Experiments). The GSinCE procedure is based on a two-stage group screening ap- proach, in which groups of inputs are investigated in the first stage and then inputs within only those groups identified as active at the first stage are investigated indi- vidually at the second stage. Two-stage designs with desirable properties are con- structed to implement the procedure.
    [Show full text]
  • Computational Physics 1 Home Page Title Page
    Computational Physics 1 Home Page Title Page Module PH3707 Contents JJ II 10 Credits J I Page1 of 85 Go Back by Roger Stewart Full Screen Version date: Monday, 7 July 2008 Close Quit opttoa Physics Computational opttoa Physics Computational opttoa Physics Computational opttoa Physics Computational Home Page opttoa Physics Computational Physics Computational Title Page Contents opttoa Physics Computational Physics Computational opttoa Physics– Computational JJ II Computational Physics– opttoa Physics– Computational Computational Physics– opttoa Physics– Computational Computational Physics– opttoa Physics– Computational Computational Physics– J I Computational Physics Physics– Computational Computational Physics– Computational Physics opttoa Physics– Computational Computational Physics– Page2 of 85 opttoa Physics– Computational Computational Physics– opttoa Physics– Computational Computational Physics– opttoa Physics– Computational Computational Physics– Go Back Computational Physics Computational Physics Full Screen Close Computational Physics Computational Physics Quit Computational Physics Computational Physics ComputationalComputational Physics Physics Copyright c 2003 Dennis Dunn. Home Page Title Page Contents JJ II J I Page3 of 85 Go Back Full Screen Close Quit Contents Home Page Title Page Contents Contents 4 JJ II INTRODUCTION TO THE MODULE5 J I 1 ANALYSIS OF WAVEFORMS 21 Page4 of 85 1.1 Objectives 22 Go Back 1.2 Fourier Analysis 23 1.2.1 Why bother? 24 Full Screen 1.2.2 Aperiodic functions 25 Close 1.3 The Numerical Methods 26 Quit
    [Show full text]
  • Uncertainty Quantification of Stochastic Simulation for Black-Box Computer Experiments
    Methodol Comput Appl Probab https://doi.org/10.1007/s11009-017-9599-7 Uncertainty Quantification of Stochastic Simulation for Black-box Computer Experiments Youngjun Choe1 · Henry Lam2 · Eunshin Byon3 Received: 23 September 2016 / Revised: 15 October 2017 / Accepted: 18 October 2017 © Springer Science+Business Media, LLC 2017 Abstract Stochastic simulations applied to black-box computer experiments are becoming more widely used to evaluate the reliability of systems. Yet, the reliability evaluation or computer experiments involving many replications of simulations can take significant com- putational resources as simulators become more realistic. To speed up, importance sampling coupled with near-optimal sampling allocation for these experiments is recently proposed to efficiently estimate the probability associated with the stochastic system output. In this study, we establish the central limit theorem for the probability estimator from such pro- cedure and construct an asymptotically valid confidence interval to quantify estimation uncertainty. We apply the proposed approach to a numerical example and present a case study for evaluating the structural reliability of a wind turbine. Keywords Central limit theorem · Confidence interval · Importance sampling · Monte Carlo simulation · Variance reduction Mathematics Subject Classification (2010) 65C05 · 68U20 · 62F12 Electronic supplementary material The online version of this article (https://doi.org/10.1007/s11009-017-9599-7) contains supplementary material, which is available to authorized
    [Show full text]
  • Replication Or Exploration? Sequential Design for Stochastic Simulation Experiments
    Replication or exploration? Sequential design for stochastic simulation experiments Micka¨elBinois∗ Jiangeng Huangy Robert B. Gramacyy Mike Ludkovskiz January 28, 2019 Abstract We investigate the merits of replication, and provide methods for optimal design (including replicates), with the goal of obtaining globally accurate emulation of noisy computer simulation experiments. We first show that replication can be beneficial from both design and computational perspectives, in the context of Gaussian process surrogate modeling. We then develop a lookahead based sequential design scheme that can determine if a new run should be at an existing input location (i.e., replicate) or at a new one (explore). When paired with a newly developed heteroskedastic Gaus- sian process model, our dynamic design scheme facilitates learning of signal and noise relationships which can vary throughout the input space. We show that it does so efficiently, on both computational and statistical grounds. In addition to illustrative synthetic examples, we demonstrate performance on two challenging real-data simula- tion experiments, from inventory management and epidemiology. Keywords: computer experiment, Gaussian process, surrogate model, input-dependent noise, replicated observations, lookahead 1 Introduction Historically, design and analysis of computer experiments focused on deterministic solvers from the physical sciences via Gaussian process (GP) interpolation (Sacks et al., 1989). But arXiv:1710.03206v2 [stat.ME] 25 Jan 2019 nowadays computer modeling is common
    [Show full text]