An adaptive CP method for TSP solving

Nicolas Isoart (1) , Jean-Charles Régin (1)

1 I3S - Laboratoire d'Informatique, Signaux, et Systèmes de Sophia Antipolis

M. Sellmann showed that CP-based Lagrangian relaxation gave good results but the interactions between the two techniques were quite dicult to understand. There are two main reasons for this: the best multipliers do not lead to the best ltering and each ltering disrupts the Lagrangian multiplier problem (LMP) to be solved. As the resolution of the TSP in CP is mainly based on a Lagrangian relaxation, we propose to study in detail these interactions for this particular problem. This article experimentally conrms the above statements and shows that it is very dicult to establish any relationship between ltering and the method used to solve the LMP in practice. Thus, it seems very dicult to select a priori the best method suited for a given instance. We propose to use a multi-armed bandit algorithm to nd the best possible method to use. The experimental results show the advantages of our approach.

Domaines: Informatique [cs]; Intelligence artificielle [cs.AI]

Nicolas Isoart, Jean-Charles Régin. An adaptive CP method for TSP solving. [Research Report] Université Côte d'Azur. 2019. ⟨hal-02165374⟩