Manipulating Models in Artificial Intelligence and Operations Research

点击翻译按钮获取中文摘要

Algorithmic decision making is widespread from the execution of supply chains to robot planning. My program will research the creation, learning, evaluation, hybridization, application, and orchestration of the models and solvers of AI and OR to push problem solving power beyond the current state of the art and to develop a fundamental mathematical and empirical understanding of such manipulation. I will work at the intersection of AI and OR within three themes. In each, I will bring together theory and algorithms from disparate subfields of computer science and mathematics to provide new perspectives and spur novel avenues of inquiry. Theme 1: Model Hierarchies in AI and OR. Complex problems are often attacked by breaking them into pieces. A problem is represented as a hierarchy of models where solving a higher level model defines a subspace at a lower level and solving a lower level model triggers refinement at the higher level. This approach, called abstraction in AI and decomposition in OR, appears in several variants (e.g., logic-based Benders decomposition, SAT Modulo Theory). I will seek a synthesis across domains, develop empirically driven theories, and create automated techniques for the dynamic creation and modification of model hierarchies. Theme 2: Model Manipulations for Sequence Generation. Modern machine learning approaches have shown substantial performance in sequence-to-sequence tasks. An ML model provides a probability distribution over the next token as the sequence is searched over. AI planning and combinatorial optimization also often generate sequences, guided by heuristic functions. This theme will develop and hybridize this intersection by investigating learned models in AI planning and combinatorial optimization, the combination of learned and built models, and the use and analysis of search techniques from AI planning, heuristic search, and combinatorial optimization for ML model decoding. Theme 3: Coordinating Algorithmic Decision Making Systems. Modern AI and OR systems require applications to be modeled, results to be interpreted, and the solvers to be coordinated as problems and contexts change. These functions have traditionally been done by highly trained humans. This theme will investigate the combination of cognitive models from AI, mixed initiative user interface design, and robot control architectures to develop an Optimization Agent to automate these model manipulations. The work in my Discovery program is aimed at conceptual level advances that can be implemented across multiple applications. These themes are informed and motivated by real world applications including advanced scheduling and logistics; robot task allocation; and scheduling for large-scale data centres. As part of my Discovery grant program, I am also proposing an application-based collaborative project to design, optimize, and operate the multi-modal transportation network of remote communities in Canada's North.