Bayesian Experimental Design with Active Learning Algorithms
Abstract
Modern societies have reached such high levels of sophistication that real-world systems (e.g.,social networks, financial markets, biological systems, artificial-intelligence algorithms) havebecome far too intricate to be designed and optimized using traditional techniques. In fact, so littleis known about the internal workings of those systems that practitioners have no choice but to treatthem as black boxes, for which a high evaluation cost adds to the issue of opacity, makingoptimization and design a daunting endeavor. This conundrum has led to a number of studies whichresulted in the emergence of iterative Bayesian algorithms, currently viewed as the state of the art.A key component in iterative Bayesian algorithms lies in the choice of the criterion that guides theexploration of the input space. These criteria come in many shapes and forms and enjoy variouslevels of popularity depending mostly on efficiency, computational complexity, implementability,and versatility. The two key issues are the ability to identify the most relevant regions in theinput space, which can only be done by incorporating information about the output space; andtractability in high dimensions, which de facto excludes any sort of Monte Carlo sampling. Wepropose to leverage ideas from the theory of importance sampling in order to design novel criteriathat address the above key issues. The central idea is to incorporate a bias in the search algorithmthat identifies and focuses on the most relevant features of the black box. We propose the study ofthe theoretical and numerical study of the new criteria in a variety of real-world applications thatpose significant challenges, involving both engineering and physical systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112357
Entities
People
- Themistoklis Sapsis
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy