Bayesian Optimization With Parallel Function Evaluations And Multiple Information Sources

Bayesian optimization, a framework for global optimization of expensive-to-evaluate functions, has recently gained popularity in machine learning and global optimization because it can find good feasible points with few function evaluations. In this dissertation, we present novel Bayesian optimization algorithms for problems with parallel function evaluations and multiple information sources, for use in machine learning, biochemistry, and aerospace engineering applications. First, we present a novel algorithm that extends expected improvement, a widely-used Bayesian optimization algorithm that evaluates one point at a time, to settings with parallel function evaluations. This algorithm is based on a new efficient solution method for finding the Bayes-optimal set of points to evaluate next in the context of parallel Bayesian optimization. The author implemented this algorithm in an open source software package co-developed with engineers at Yelp, which was used by Yelp and Netflix for automatic tuning of hyperparameters in machine learning algorithms, and for choosing parameters in online content delivery systems based on evaluations in A/B tests on live traffic. Second, we present a novel parallel Bayesian optimization algorithm with a worst-case approximation guarantee applied to peptide optimization in biochemistry, where we face a large collection of peptides with unknown fitness prior to experimentation, and our goal is to identify peptides with a high score using a small number of experiments. High scoring peptides can be used for biolabeling, targeted drug delivery, and self-assembly of metamaterials. This problem has two novelties: first, unlike traditional Bayesian optimization, where the objective function has a continuous domain and real-valued output well-modeled by a Gaussian Process, this problem has a discrete domain, and involves binary output not well-modeled by a Gaussian process; second, it uses hundreds of parallel function evaluations, which is a level of parallelism too large to be approached with other previously-proposed parallel Bayesian optimization methods. Third, we present a novel Bayesian optimization algorithm for problems in which there are multiple methods or "information sources" for evaluating the objective function, each with its own bias, noise and cost of evaluation. For example, in aerospace engineering, to evaluate an aircraft wing design, different computational models may simulate performance. Our algorithm explores the correlation and model discrepancy of each information source, and optimally chooses the information source to evaluate next and the point at which to evaluate it. We describe how this algorithm can be used in general multi information source optimization problems, and also how a related algorithm can be used in "warm start" problems, where we have results from previous optimizations of closely related objective functions, and we wish to leverage these results to more quickly optimize a new objective function.

This volume brings together the main results in the field of Bayesian Optimization (BO), focusing on the last ten years and showing how, on the basic framework, new methods have been specialized to solve emerging problems from machine learning, artificial intelligence, and system optimization. It also analyzes the software resources available for BO and a few selected application areas. Some areas for which new results are shown include constrained optimization, safe optimization, and applied mathematics, specifically BO's use in solving difficult nonlinear mixed integer problems. The book will help bring readers to a full understanding of the basic Bayesian Optimization framework and gain an appreciation of its potential for emerging application areas. It will be of particular interest to the data science, computer science, optimization, and engineering communities.

This open access book presents the first comprehensive overview of general methods in Automated Machine Learning (AutoML), collects descriptions of existing systems based on these methods, and discusses the first series of international challenges of AutoML systems. The recent success of commercial ML applications and the rapid growth of the field has created a high demand for off-the-shelf ML methods that can be used easily and without expert knowledge. However, many of the recent machine learning successes crucially rely on human experts, who manually select appropriate ML architectures (deep learning architectures or more traditional ML workflows) and their hyperparameters. To overcome this problem, the field of AutoML targets a progressive automation of machine learning, based on principles from optimization and machine learning itself. This book serves as a point of entry into this quickly-developing field for researchers and advanced students alike, as well as providing a reference for practitioners aiming to use AutoML in their work.

Bayesian optimization helps pinpoint the best configuration for your machine learning models with speed and accuracy. Put its advanced techniques into practice with this hands-on guide. In Bayesian Optimization in Action you will learn how to: Train Gaussian processes on both sparse and large data sets Combine Gaussian processes with deep neural networks to make them flexible and expressive Find the most successful strategies for hyperparameter tuning Navigate a search space and identify high-performing regions Apply Bayesian optimization to cost-constrained, multi-objective, and preference optimization Implement Bayesian optimization with PyTorch, GPyTorch, and BoTorch Bayesian Optimization in Action shows you how to optimize hyperparameter tuning, A/B testing, and other aspects of the machine learning process by applying cutting-edge Bayesian techniques. Using clear language, illustrations, and concrete examples, this book proves that Bayesian optimization doesn’t have to be difficult! You’ll get in-depth insights into how Bayesian optimization works and learn how to implement it with cutting-edge Python libraries. The book’s easy-to-reuse code samples let you hit the ground running by plugging them straight into your own projects. Forewords by Luis Serrano and David Sweet. About the technology In machine learning, optimization is about achieving the best predictions—shortest delivery routes, perfect price points, most accurate recommendations—in the fewest number of steps. Bayesian optimization uses the mathematics of probability to fine-tune ML functions, algorithms, and hyperparameters efficiently when traditional methods are too slow or expensive. About the book Bayesian Optimization in Action teaches you how to create efficient machine learning processes using a Bayesian approach. In it, you’ll explore practical techniques for training large datasets, hyperparameter tuning, and navigating complex search spaces. This interesting book includes engaging illustrations and fun examples like perfecting coffee sweetness, predicting weather, and even debunking psychic claims. You’ll learn how to navigate multi-objective scenarios, account for decision costs, and tackle pairwise comparisons. What's inside Gaussian processes for sparse and large datasets Strategies for hyperparameter tuning Identify high-performing regions Examples in PyTorch, GPyTorch, and BoTorch About the reader For machine learning practitioners who are confident in math and statistics. About the author Quan Nguyen is a research assistant at Washington University in St. Louis. He writes for the Python Software Foundation and has authored several books on Python programming. Table of Contents 1 Introduction to Bayesian optimization 2 Gaussian processes as distributions over functions 3 Customizing a Gaussian process with the mean and covariance functions 4 Refining the best result with improvement-based policies 5 Exploring the search space with bandit-style policies 6 Leveraging information theory with entropy-based policies 7 Maximizing throughput with batch optimization 8 Satisfying extra constraints with constrained optimization 9 Balancing utility and cost with multifidelity optimization 10 Learning from pairwise comparisons with preference optimization 11 Optimizing multiple objectives at the same time 12 Scaling Gaussian processes to large datasets 13 Combining Gaussian processes with neural networks

Many science, engineering, and design optimization problems require balancing the trade-offs between several conflicting objectives. The objectives are often blackbox functions whose evaluation requires time-consuming and costly experiments. Multi-objective Bayesian optimization can be used to automate the process of discovering the set of optimal solutions, called Pareto-optimal, while minimizing the number of performed evaluations. To further reduce the evaluation time in the optimization process, testing of several samples in parallel can be deployed. We propose DGEMO, a novel multi-objective Bayesian optimization algorithm that iteratively selects the best batch of samples to be evaluated in parallel. Our algorithm approximates and analyzes a piecewise-continuous Pareto set representation, which allows us to introduce a batch selection strategy that optimizes for both hypervolume improvement and diversity of selected samples in order to efficiently advance promising regions of the Pareto front. Experiments on both synthetic test functions and real-world benchmark problems show that our algorithm predominantly outperforms relevant state-of-the-art methods. The code is available at https://github.com/yunshengtian/DGEMO. In addition, we present AutoOED, an Optimal Experiment Design platform that implements several multi-objective Bayesian optimization algorithms with state-of-the-art performance including DGEMO with an intuitive graphical user interface (GUI). AutoOED is open-source and written in Python. The codebase is modular, facilitating extensions and tailoring the code, serving as a testbed for machine learning researchers to easily develop and evaluate their own multi-objective Bayesian optimization algorithms. Furthermore, a distributed system is integrated to enable parallelized experimental evaluations by independent workers in remote locations. The platform is available at https://autooed.org.

This book is dedicated to the 70th birthday of Professor J. Mockus, whose scientific interests include theory and applications of global and discrete optimization, and stochastic programming. The papers for the book were selected because they relate to these topics and also satisfy the criterion of theoretical soundness combined with practical applicability. In addition, the methods for statistical analysis of extremal problems are covered. Although statistical approach to global and discrete optimization is emphasized, applications to optimal design and to mathematical finance are also presented. The results of some subjects (e.g., statistical models based on one-dimensional global optimization) are summarized and the prospects for new developments are justified. Audience: Practitioners, graduate students in mathematics, statistics, computer science and engineering.

Bayesian Optimization in Action teaches you how to build Bayesian Optimisation systems from the ground up. This book transforms state-of-the-art research into usable techniques you can easily put into practice. With a range of illustrations, and concrete examples, this book proves that Bayesian Optimisation doesn't have to be difficult!

This book addresses the current status, challenges and future directions of data-driven materials discovery and design. It presents the analysis and learning from data as a key theme in many science and cyber related applications. The challenging open questions as well as future directions in the application of data science to materials problems are sketched. Computational and experimental facilities today generate vast amounts of data at an unprecedented rate. The book gives guidance to discover new knowledge that enables materials innovation to address grand challenges in energy, environment and security, the clearer link needed between the data from these facilities and the theory and underlying science. The role of inference and optimization methods in distilling the data and constraining predictions using insights and results from theory is key to achieving the desired goals of real time analysis and feedback. Thus, the importance of this book lies in emphasizing that the full value of knowledge driven discovery using data can only be realized by integrating statistical and information sciences with materials science, which is increasingly dependent on high throughput and large scale computational and experimental data gathering efforts. This is especially the case as we enter a new era of big data in materials science with the planning of future experimental facilities such as the Linac Coherent Light Source at Stanford (LCLS-II), the European X-ray Free Electron Laser (EXFEL) and MaRIE (Matter Radiation in Extremes), the signature concept facility from Los Alamos National Laboratory. These facilities are expected to generate hundreds of terabytes to several petabytes of in situ spatially and temporally resolved data per sample. The questions that then arise include how we can learn from the data to accelerate the processing and analysis of reconstructed microstructure, rapidly map spatially resolved properties from high throughput data, devise diagnostics for pattern detection, and guide experiments towards desired targeted properties. The authors are an interdisciplinary group of leading experts who bring the excitement of the nascent and rapidly emerging field of materials informatics to the reader.

A statistical view of uncertainty in expert systems. Knowledge, decision making, and uncertainty. Conceptual clustering and its relation to numerical taxonomy. Learning rates in supervised and unsupervised intelligent systems. Pinpoint good hypotheses with heuristics. Artificial intelligence approaches in statistics. REX review. Representing statistical computations: toward a deeper understanding. Student phase 1: a report on work in progress. Representing statistical knowledge for expert data analysis systems. Environments for supporting statistical strategy. Use of psychometric tools for knowledge acquisition: a case study. The analysis phase in development of knowledge based systems. Implementation and study of statistical strategy. Patterns in statisticalstrategy. A DIY guide to statistical strategy. An alphabet for statistician's expert systems.

This two-volume set LNCS 12269 and LNCS 12270 constitutes the refereed proceedings of the 16th International Conference on Parallel Problem Solving from Nature, PPSN 2020, held in Leiden, The Netherlands, in September 2020. The 99 revised full papers were carefully reviewed and selected from 268 submissions. The topics cover classical subjects such as automated algorithm selection and configuration; Bayesian- and surrogate-assisted optimization; benchmarking and performance measures; combinatorial optimization; connection between nature-inspired optimization and artificial intelligence; genetic and evolutionary algorithms; genetic programming; landscape analysis; multiobjective optimization; real-world applications; reinforcement learning; and theoretical aspects of nature-inspired optimization.