Collaborative Research: RI: Small: End-to-end Learning of Fair and Explainable Schedules for Court Systems

美国法院系统是一个庞大而复杂的社会技术系统，每年处理数百万起刑事案件。然而，目前的审前安排过程中，令人震惊的是五分之一的被告错过了开庭日期。这给作为一个机构的司法机构带来了高昂的成本，对于就业情况不稳定、照顾他人的责任或缺乏前往法院的交通工具的被告来说，可能尤其有害。这些不同的影响具有深远的负面影响。为了解决这些问题，本项目研究了公平和可解释的学习调度，这是一种紧密结合机器学习、约束优化和知识表示的新方法，用于学习具有可证明公平性保证的调度，并使神经符号推理能够提供有意义和可细化的解释。拟议的研究将开发新的工具，以确保审前安排可以减少缺席并平等对待所有被告，从而具有显著的社会效益。从科学的角度来看，该项目将开发新一代综合学习和优化工具以及解释工具，以实现更公平和更公平的时间表的潜力。提出的公平和可解释的学习到时间表将在几个领域做出重要贡献，包括：(1)使深度学习系统能够处理表示时间表的组合结构；(2)开发端到端的训练过程，将约束优化集成到学习管道中；(3)在学习过程中提供用户指定的公平概念的满意度保证；(4)开发神经符号方法来提供关于时间表和公平性质的解释；(5)将学习和基于逻辑的推理相结合，在适当的抽象层次向用户提供个性化的解释；以及(6)为公平的审前法庭安排开发新的数据集。这一裁决反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The American Court system is a large and complex socio-technical system that handles millions of criminal cases every year. However, the current pretrial scheduling process is plagued by a staggering one in five defendants missing court dates. This imposes high costs on the judiciary as an institution, and can be particularly harmful to defendants who have insecure employment situations, care-giving responsibilities, or lack transportation to court. These disparate impacts have profound negative effects. To address these issues, this project investigates Fair and Explainable Learning to Schedule, a novel approach that tightly integrates machine learning, constrained optimization, and knowledge representation to learn schedules with certifiable fairness guarantees and enable neuro-symbolic reasoning to provide meaningful and refinable explanations. The proposed research will develop new tools to ensure that pretrial scheduling can decrease nonappearance and be fair to all defendants equally and has thus the potential to have significant societal benefits.From a scientific standpoint, this project will develop a new generation of integrated learning and optimization tools as well as explanation tools to realize the potential of fairer and more equitable schedules. The proposed Fair and Explainable Learning to Schedule will make key contributions in several areas, including: (1) enabling deep learning systems to handle combinatorial structures to represent schedules; (2) developing end-to-end training procedures that integrate constrained optimization within a learning pipeline; (3) providing guarantees on the satisfaction of user-specified fairness notions in the learning process; (4) developing neuro-symbolic approaches to provide explanations about scheduling and fairness properties; (5) integrating learning and logic-based reasoning to provide personalized explanations at appropriate abstraction levels to users; and (6) developing new datasets for fair pretrial court scheduling.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.