Résumé en anglais

Multiobjective heuristic search in state space graphs generally aims at determining theexact set of Pareto-optimal cost vectors associated to solution paths or an approximationof this Pareto set. When a preference model is available under the form of scalarizingfunction f, the search effort merely focuses on the direct determination of an f-optimaltradeoff within the Pareto-set. The solution algorithms proposed in this area rely on uni-directional approaches that develop a front of Pareto-optimal labels attached to subpathsfrom the initial node s to the goal nodes. When there is a single goal node t, explicitlyknown, and when search operators are reversible, a bidirectional search approach is possi-ble, interleaving and interacting searches, forward from s and backward from t. Althoughthe respective advantages of unidirectional approaches versus bidirectionnal approacheshave been widely discussed in the singleobjective case, they have not been investigated inthe setting of multiobjective search. The aim of this paper is to propose a first implemen-tation of multiobjective bidirectional heuristic search and to empirically assess his valuecompared to unidirectional search. Our comparison concerns both the determination ofthe entire Pareto set and the determination of OWA-optimal tradeoffs within the Paretoset. We provide results of numerical tests performed on random instances of path-planningproblems, both in terms of number of nodes expanded and in terms of computation times.