MO-ParamILS: A Multi-objective Automatic Algorithm Configuration Framework

Abstract

International audienceAutomated algorithm configuration procedures play an increasingly important role in the development and application of algorithms for a wide range of computationally challenging problems. Until very recently, these configuration procedures were limited to optimising a single performance objective, such as the running time or solution quality achieved by the algorithm being configured. However, in many applications there is more than one performance objective of interest. This gives rise to the multi-objective automatic algorithm configuration problem, which involves finding a Pareto set of configurations of a given target algorithm that characterises trade-offs between multiple performance objectives. In this work, we introduce MO-ParamILS, a multi-objective extension of the state-of-the-art single-objective algorithm configuration framework ParamILS, and demonstrate that it produces good results on several challenging bi-objective algorithm configuration scenarios compared to a base-line obtained from using a state-of-the-art single-objective algorithm configurator

Automatically Configuring Multi-objective Local Search Using Multi-objective Optimisation

“…In principle, these and other single-objective AAC procedures can easily be applied to multi-objective optimisation (MOO) target algorithms, using a single performance indicator or an aggregation of several indicators (e.g., using the hypervolume of normalised indicators [1]). A conceptually attractive alternative is to directly optimise multiple performance indicators, requiring a multi-objective configurator, such as the recently proposed MO-ParamILS [2], an extension of the original, single-objective ParamILS configuration framework. To the best of our knowledge, multi-objective configurators have so far only been applied to singleobjective target algorithms -for example, for simultaneously optimising solution quality and running time of single-objective optimisation algorithms.…”

“…We evaluated three configuration approaches: HV ||∆, a multi-objective approach, in which both hypervolume and ∆ spread are minimised, using MO-ParamILS [2]; HV , a single-objective approach, in which only hypervolume is minimised; and HV +∆, a single-objective approach, in which we minimise a weighted sum of hypervolume and ∆-spread. For both single-objective approaches, we used single-objective ParamILS [11], as implemented in our MO-ParamILS framework.…”

“…Thus, it is becoming increasingly common practice to use automatic algorithm configuration (AAC) procedures, such as irace [14], ParamILS [11] or SMAC [10]. While these configurators optimise a single performance metric only, such as solution quality or running time of a given target algorithm, recently, multi-objective AAC procedures, such as MO-ParamILS [2], have become available and shown to be effective for optimising multiple performance metric simultaneously, using a multi-objective approach.…”

Automatic design of multi-objective local search algorithms

“…Recently, multi-objective con gurators such as SPRINT-race [23] or MOParamILS [2] have shown the possibility and the bene ts of using multiple performance indicators. It has also been shown that, on multi-objective AAC scenario optimising multiple multi-objective performance indicators of multi-objective target algorithms, using a multi-objective con gurator should be preferred to using a singleobjective con gurator with an aggregation of the performance indicators [3].…”

“…en, we investigate the performance of an AAC procedure in comparison to an exhaustive analysis of all feasible MOLS con gurations. We use MO-ParamILS [2], a multi-objective AAC con gurator, in order to account for the fundamental multiobjective nature of MOLS algorithms.…”