New Frontiers in Adaptive Experimental Design for Multi-Objective Optimization

Syrine Belakaria

doi:10.7273/000005366

The problem of adaptively selecting a sequence of experiments to achieve a goal (aka adaptive experimental design) arises in many real-world settings. A canonical example is the active learning paradigm where we need to iteratively collect labeled data to build predictors with high accuracy. Motivated by the challenges faced by scientists and engineers, this dissertation studies adaptive experimental design algorithms for the purpose of solving a large-class of multi-objective optimization (MOO) problems. Such MOO problems enable many science and engineering applications including drug design, protein engineering, design of materials, and hardware design. For example, in drug design optimization, we need to find drugs that trade-off effectiveness, safety, and cost by performing expensive experiments to evaluate each candidate drug. Similarly, in hardware design optimization, we need to find the designs that trade-off performance, energy, and area using expensive computational simulations to mimic the real hardware. We have the ability to evaluate any candidate input according to the target objectives by performing a costly experiment, where the cost is measured by the resources consumed by the experiment (physical or computational). Our overall goal is to approximate the optimal Pareto set of solutions by minimizing the total resource cost of conducted experiments. The key challenge is how to select the sequence of experiments under uncertainty. This dissertation develops a suite of novel reasoning algorithms based on the principles of information gain per unit resource cost and uncertainty reduction for adaptive experimental design to solve MOO problems. We appropriately instantiate these principles to derive efficient algorithms for the following MOO problem settings, most of which are studied for the first time: 1) The most basic single-fidelity setting, where experiments are expensive and accurate, and we can conduct a single experiment in each iteration; 2) The batch setting where a batch of experiments can be conducted in parallel to accelerate the search process; 3) The constrained setting where we cannot evaluate constraints to identify feasible inputs without performing experiments; 4) The discrete multi-fidelity setting where experiments can vary in the amount of resources consumed and their evaluation accuracy; and 5) The continuous-fidelity setting, where continuous function approximations result in a huge space of experiments. 6) The budget-aware setting where a limited resource budget constraint is enforced requiring us to take a planning approach. Experiments on synthetic and real-world benchmarks from a diverse set of engineering and industrial domains demonstrate that our algorithms significantly improve resource efficiency over prior methods to produce high-quality Pareto solutions.

New Frontiers in Adaptive Experimental Design for Multi-Objective Optimization

Files and links (1)

Abstract

Metrics

Details