Simple Rules for Low-Knowledge Algorithm Selection

This paper addresses the question of selecting an algorithm from a predefined set that will have the best performance on a scheduling problem instance. Our goal is to reduce the expertise needed to apply constraint technology. Therefore, we investigate simple rules that make predictions based on limited problem instance knowledge. Our results indicate that it is possible to achieve superior performance over choosing the algorithm that performs best on average on the problem set. The results hold over a variety of different run lengths and on different types of scheduling problems and algorithms. We argue that low-knowledge approaches are important in reducing expertise required to exploit optimization technology.