Instance-Specific Parameter Tuning for Meta-Heuristics

Meta-heuristics are of significant interest to decision-makers due to the capability of finding good solutions for complex problems within a reasonable amount of computational time. These methods are further known to perform according to how their algorithm-specific parameters are set. As most practitioners aim for an off-the-shelf approach when using meta-heuristics, they require an easy applicable strategy to calibrate its parameters and use it. This chapter addresses the so-called Parameter Setting Problem (PSP) and presents new developments for the Instance-specific Parameter Tuning Strategy (IPTS). The IPTS presented only requires the end user to specify its preference regarding the trade-off between running time and solution quality by setting one parameter p (0 ≤ p ≤1), and automatically returns a good set of algorithm-specific parameter values for each individual instance based on the calculation of a set of problem instance characteristics. The IPTS does not require any modification of the particular meta-heuristic being used. It aims to combine advantages of the Parameter Tuning Strategy (PTS) and the Parameter Control Strategy (PCS), the two major approaches to the PSP. The chapter outlines the advantages of an IPTS and shows in more detail two ways in which an IPTS can be designed. The first design approach requires expert-based knowledge of the meta-heuristic’s performance in relation to the problem at hand. The second, automated approach does not require explicit knowledge of the metaheuristic used. Both designs use a fuzzy logic system to obtain parameter values. Results are presented for an IPTS designed to solve instances of the Travelling Salesman Problem (TSP) with the meta-heuristic Guided Local Search (GLS). Jana Ries University of Portsmouth, UK Patrick Beullens University of Southampton, UK Yang Wang Beijing University of Technology, China