System design space exploration requires searching for inputs that achieve the desired performance. When the system evaluation is expensive, the designer needs to minimize the number of system evaluation calls. Real-word system design often involves repetitive design to meet different customer needs. The designer may wish to dynamically drive a system from one specification to another without starting the search from scratch. Furthermore, when system evaluation is expensive, the designer may wish to gauge how much confidence to place in the solution suggested by the design tool and determine whether to evaluate a lower confidence solution with the possibility of getting a larger improvement in performance or vice versa. Finally, real-world data often have errors. We need a mechanism that detects and locates/corrects errors in a data instance to enhance data quality.

This thesis addresses the above challenges. First, we introduce a sample-efficient system design framework called ASSENT. ASSENT uses a two-step methodology. In Step 1, we use a genetic algorithm (GA) to find a coarse system design. The design is improved in Step 2 using a neural network (NN) verifier. Next, we present INFORM, which uses a hybrid GA in Step 1. We inject candidate solutions into the GA utilizing a combination of three inverse design methods based on an NN verifier, a Gaussian Mixture Model (GMM), and an NN. In Step 2, we use inverse design methods to improve the quality of the reference solution. Finally, we present REPAIRS: an explainable decision framework to complete/optimize a partially-specified system, and correct errors in a data instance. It uses a GMM to learn the joint distribution of system input and response. We use the learned model to complete a partially-specified system where only a subset of the component values and/or the system response is specified. When the system response exhibits multiple modes, REPAIRS determines the combinations of input values that correspond to the several modes. We also use REPAIRS to verify the integrity of a given data instance. When the integrity check fails, we provide a mechanism to identify and correct the error.

Adviser: Niraj Jha

Zoom Meeting Link: https://princeton.zoom.us/j/93117303397