In Situ Reconfiguration of Heuristic Search on a Problem Instance Basis

The time it takes a program to solve a particular problem depends heavily upon the choice of problem solving method, the data representation, heuristics etc. The specific choices can have a dramatic impact on performance. Our objective is to design a problem solving method which dynamically adapts its search configuration in order to speed up finding a solution. Previous approaches have used performance data gathered on past problem solving episodes to predict performance of different problem solving configurations on future problem instances. These approaches prediction quality depend on the future problem distribution being known. The main novelty of our approach is that we reconfigure the search based solely on performance data gathered while solving the current problem instance. As means to this end we use a formula based on these design decisions and problem characteristics which predicts how long the problem will run until a solution is found.