Park-and-Ride Facility Location Selection under Nested Logit Demand Function

TL;DR

This paper extends the park-and-ride facility location problem by incorporating a nested logit demand model to better reflect real-world choice behavior, and develops efficient algorithms to solve large-scale instances exactly.

Contribution

It introduces a nested logit demand model for P&R location problems and proposes two computational methods, neighborhood search and randomized rounding, for large-scale optimization.

Findings

Methods solve 1,000 medium-scale instances rapidly.

Achieve exact optimality 10,000 times faster than previous MNL-based methods.

Highlight differences between MNL and NL models in P&R context.

Abstract

Park-and-ride facilities are car parks where users can transfer to public transportation. Commuters can use P&R facilities or choose to travel by car to their destinations, and individual choice behavior is assumed to follow a logit model. The P&R facility location problem identifies locations for a fixed number of P&R facilities from among potential locations such that the number of users of the P&R facilities is maximized. This problem has previously been formalized under a multinomial logit demand function. However, as it imposes the strong condition of the independence of irrelevant alternatives, the MNL model is unable to represent the real-world P&R facility location problem exactly. Respecting the nested structure of individual choice behavior, we generalize the MNL model to a nested logit model and develop two computational methods -- neighborhood search and randomized rounding -- to solve large-scale P&R facility location problems under the NL demand function. The neighborhood search method first finds one feasible solution and then improves the feasible solution to the next one along an edge of the polyhedron whose vertices are the feasible solutions. The neighborhood search method is randomized to create an adaptive randomized rounding procedure. Computational experiments verified that the computational methods were able to solve the nonlinear optimization problem under the NL demand function on 1,000 medium-scale instances to exact optimality rapidly. Specifically, the methods were verified to solve the MNL model to exact optimality 10,000 times faster than the mixed integer linear programming formulation described in the literature. We performed additional computational experiments to assess the performance of our computational methods on a variety of large-scale instances. Our computational analysis also elucidates the difference between the MNL and NL models.