Research Projects for Graduate Students Exploiting Randomisation in Search

This document outlines a number of projects for which I am seeking to recruit graduate students. Of the projects detailed below, the rst three will be conducted in collaboration with Toby Walsh. In general I have a range of research interests to do with knowledge representation , automated reasoning and machine learning. Results are applied to problems in planning, natural language processing, logic programming, constraint satisfaction and optimisation, inductive learning, probabilistic reasoning, and veriication. Much excitement has been generated by the recent success of stochastic local search procedures at nding solutions to large, very hard search problems. Local search procedures move through a space of potential solutions to a problem guided by a heuristic that at each point chooses what appears to be the best move from among the alternatives. It has been found that such a procedure performs much better if at each step, with a predetermineed probability (called the noise setting), it ignores the heuristic and randomly chooses a move from among the set of alternatives SKC94]. Furthermore, our research FP99] has shown that the performance of these stochastic local search procedures varies enormously across the range of noise settings, and the optimal noise setting varies enormously from problem instance to problem instance. This project addresses the question of how we can get the most out of stochastic local search procedures in spite of this great variation of performance. One approach that will be taken is to develop a method for predicting the optimal noise setting for any given problem instance. Another approach that will be taken is to develop a stochastic local search proce