Parameter tuning is recognized today as a crucial ingredient when tackling anoptimization problem. Several meta-optimization methods have been proposed tofind the best parameter set for a given optimization algorithm and (set of)problem instances. When the objective of the optimization is some scalarquality of the solution given by the target algorithm, this quality is alsoused as the basis for the quality of parameter sets. But in the case ofmulti-objective optimization by aggregation, the set of solutions is given byseveral single-objective runs with different weights on the objectives, and itturns out that the hypervolume of the final population of each single-objectiverun might be a better indicator of the global performance of the aggregationmethod than the best fitness in its population. This paper discusses this issueon a case study in multi-objective temporal planning using the evolutionaryplanner DaE-YAHSP and the meta-optimizer ParamILS. The results clearly show howParamILS makes a difference between both approaches, and demonstrate thatindeed, in this context, using the hypervolume indicator as ParamILS target isthe best choice. Other issues pertaining to parameter tuning in the proposedcontext are also discussed.
展开▼
