EAGER: Requirements Domain Specifications for Machine-Learned Software Components

点击翻译按钮获取中文摘要

Software components are traditionally built according to a set of pre-defined specifications that define what a software component is expected to do in a given context. These specifications are traditionally gathered from stakeholders and customers during the domain-analysis and requirements-engineering phases. In contrast to deterministic software components, software components built using Machine-Learning (ML) algorithms learn their specifications from a set of collected examples rather than a set of agreed upon specifications. When the functional correctness of ML-enabled software depends only on the training data, there can be a significant gap between specification of a real-world concept and what a collected dataset represents as the concept. The goal of this research is to define the meaning of requirements satisfaction for software with machine-learning components, and to investigate methods for engineering those requirements. In this project, the investigators will formally specify partial requirements for Machine-Learned Components (MLCs) instead of allowing them to learn these specifications solely from a set of collected samples in an ad-hoc manner. The goal is thus to make machine-learning components better meet requirements by augmenting the inductive nature of ML with domain analysis, in order to characterize the extent to which the dataset contains or lacks important features that are necessary to meet requirements. This project provides a framework for formally specifying partial requirements as well as validating the presence of such specifications in the collected samples, which in essence characterizes the extent to which the dataset contains or lacks features important to learning the task. The proposal considers this problem in the context of automated driving systems where the correct definition of real-world concepts is critically important for safety reasons. For example, correct image recognition is needed to classify objects to avoid; the objective is to show that combining training by datasets with partial domain models from elicited requirements can outperform brute-force ML.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.