Searching of equilibriums in hierarchical congestion population games

TL;DR

This paper introduces a primal-dual approach for finding equilibria in hierarchical congestion population games, utilizing a multi-level convex optimization framework and a novel computational method based on a smooth Bellman-Ford technique.

Contribution

It formulates a variational principle for hierarchical congestion games and develops an efficient primal-dual method using a smooth shortest path approach for equilibrium computation.

Findings

Effective primal-dual method for hierarchical congestion games

Novel smooth Bellman-Ford based technique for graph characteristic function calculation

Reduction of equilibrium search to a multi-level convex optimization problem

Abstract

An universal primal-dual approach of description equilibriums in large class of hierarchical congestion population games is proposed. At the very core of the approach is hierarchy of enclosed to each other transport networks. In different time-scales corresponding logit dynamics on this networks is considered. This dynamics reflect restricted rationality of the agents. Searching of equilibrium configuration to the multi-level convex optimization problem is reduced (in other words variational principle for these class of the games/models is formulated). Then the dual problem is formulated. This problem has natural interpretation in turn and it is more computationally attractive then the primal one. So an effective primal-dual method for this problem is proposed. This method is based on the composite fast gradient method. Due to primal-duality the solution of the primal problem from the dual iterative process is recovered. To fulfil one iteration of this method the rather complex problem of calculation characteristic function on graph is solved. A special technique (smooth version of Bellman--Ford approach to the shortest path problem) is proposed (based on some ideas of Yu. Nesterov) that allows to do it in a cheap manner. It also should be mentioned that different tricks of small parameters methods are actively used.