Automated Algorithm Configuration and Parameter Tuning

Computationally challenging problems arise in the context of many applications, and the ability to solve these as efficiently as possible is of great practical, and often also economical importance. Examples of such problems include scheduling, timetabling, resource allocation, production planning and optimisation, computer-aided design and software verification. Many of these problems are NP-hard and considered computationally intractable, because there is no polynomial-time algorithm that can find solutions in the worst case (unless NP=P). However, by using carefully crafted heuristic techniques, it is often possible to solve practically relevant instances of these ‘intractable’ problems surprisingly effectively (see, e.g., Prasad et al, 2005; Applegate et al, 2006; Pop et al, 2002). 1 The practically observed efficacy of these heuristic mechanisms remains typically inaccessible to the analytical techniques used for proving theoretical complexity results, and therefore needs to be established empirically, on the basis of carefully designed computational experiments. In many cases, state-of-the-art performance is achieved using several heuristic mechanisms that interact in complex, non-intuitive ways. For example, a DPLL-style complete solver for SAT (a prototypical NPcomplete problem with important applications in the design of reliable softand hardware) may use different heuristics for selecting variables to be instantiated and the values first explored for these variables, as well as heuristic mechanisms for managing and using logical constraints derived from failed solution attempts. The activation, interaction and precise behaviour of those mechanisms is often controlled by parameters, and the settings of such parameters have a substantial impact on the