Research Overview
My research has focussed on the integration of constraint programming with techniques from artificial intelligence, machine learning and operations research. The goal is to develop techniques which allow constraint programming solvers to efficiently solve large scale, real world problems.
In a project invovling machine learning, we used an optimal but inefficient constraint program to generate exact labels for an previously unlabelled data set. With this labelled data, we learnt a classification heuristic which approximated the labelling. We applied this work to the problem of instruction scheduling during code generation. Our results show that this technique created a heuristic that was as good as previously hand-crafted techniques.
In a project that combined constraint programming and operations research, we showed that using enumeration and network flows can be successfully used to create more compact constraint programming models and elicity partial symmetries from the models. This technique was applied to the problem of finding the exact number of points needed by teams to either qualify for or avoid elimination from the National Hockey League Playoffs. This technique provided the first scalable technique which could solve the problem under all tie-breaking constraints.
In the area of computational social choice, we showed that the manipulation of common competitions could be manipulated by the participants. Extending this work, we used constraint programming to examine possible methods to reduce the effectiveness of cheating strategies like seeding manipulation (also known as schedule control).
In recent work, we have been examining the use of counting to improve the quality of search heuristics. The goal of this project is to create more effective general heuristics, which relieves the user of hand crafting the heuristic.
