Managing Complex Scheduling Problems with Dynamic and Hybrid Constraints.

Abstract

The task of scheduling can often be a difficult one because of the inherent complexity of real-world problems. In the field of Artificial Intelligence, many representations and algorithms have been developed to automate the scheduling process. Many state of the art scheduling systems deal with this complexity by making assumptions that simplify the algorithms, but in doing so, miss some opportunities to improve performance. Scheduling problems are temporal in nature, and so they often contain constraints that change over time. Many scheduling systems assume that the problems they are solving are all independent, and so they ignore the similarities between subsequent sets of scheduling constraints. Additionally, scheduling problems often contain a mixture of finite-domain and temporal constraints. Many of the systems that can solve problems of this type do so by creating finite-domain variables to represent the constraints, but then ignore the distinction between the different types of variables when searching for a solution. In this dissertation, I identify opportunities to improve performance by exploiting structure where it has previously been overlooked. Following this approach, I develop a set of techniques that apply to a wide variety of situations that can arise in real-world scheduling problems. First, I consider dynamic scheduling problems with constraints that change over time. To address such problems, I introduce a new representation called the Dynamic Disjunctive Temporal Problem, along with several techniques to improve both efficiency and stability when solving one. Second, I consider scheduling problems in which a mixture of finite-domain and temporal variables can interact through hybrid constraints. I introduce the Hybrid Scheduling Problem to represent such problems, and I present a set of techniques that capitalize on the distinction between variable types to improve efficiency across the problem space. Finally, I conclude by proposing several ways that the dynamic and hybrid representations and techniques can be combined. To compare many of the techniques presented throughout this dissertation in the context of structured, real-world problems, I use them to solve scheduling problems based on actual air traffic control constraints recorded from the Dallas/Fort Worth International Airport.