Vector-Value Markov Decision Process for multi-objective stochastic path planning
Résumé
The problem of path planning in stochastic environments where the shortest path is not always the best one is a challenging issue required in many real-world applications such as autonomous vehicles, robotics, logistics, etc. . . . In this paper, we consider the problem of path planning in stochastic environments where the length of the path is not the unique criterion to consider. We formalize this problem as a multi-objective decision-theoretic path planning and we transform this latter into 2VMDP (Vector-Valued Markov Decision Process). We show, then, how we can compute a policy balancing between different considered criteria. We describe different techniques that allow us to derive an optimal policy where it is hard to express the expected utilities, rewards and values with a unique numerical measure. Firstly, we examine different existing approaches based on preferences and we define notions of optimality with preferred solutions and secondly we present approaches based on egalitarian social welfare techniques. Finally, some experimental results have been developed to show the feasibility of the approach and the benefit of this approach on the single-objective techniques.
Origine : Fichiers produits par l'(les) auteur(s)