Where to Decide? Centralized vs. Distributed Vehicle Assignment for Platoon Formation

TL;DR

This paper compares centralized and distributed algorithms for vehicle-to-platoon assignment in autonomous driving, finding that a simple distributed greedy method performs nearly as well as complex centralized solutions with less complexity.

Contribution

It introduces and evaluates three approaches for vehicle-to-platoon assignment, highlighting the effectiveness of a distributed greedy heuristic in large-scale simulations.

Findings

Distributed greedy approach performs as well as centralized solver in most metrics.

Centralized greedy approach suffers from synchronization issues.

Distributed greedy requires less knowledge and complexity.

Abstract

Platooning is a promising cooperative driving application for future intelligent transportation systems. In order to assign vehicles to platoons, some algorithm for platoon formation is required. Such vehicle-to-platoon assignments have to be computed on-demand, e.g., when vehicles join or leave the freeways. In order to get best results from platooning, individual properties of involved vehicles have to be considered during the assignment computation. In this paper, we explore the computation of vehicle-to-platoon assignments as an optimization problem based on similarity between vehicles. We define the similarity and, vice versa, the deviation among vehicles based on the desired driving speed of vehicles and their position on the road. We create three approaches to solve this assignment problem: centralized solver, centralized greedy, and distributed greedy, using a Mixed Integer Programming (MIP) solver and greedy heuristics, respectively. Conceptually, the approaches differ in both knowledge about vehicles as well as methodology. We perform a large-scale simulation study using PlaFoSim to compare all approaches. While the distributed greedy approach seems to have disadvantages due to the limited local knowledge, it performs as good as the centralized solver approach across most metrics. Both outperform the centralized greedy approach, which suffers from synchronization and greedy selection effects. The centralized solver approach however assumes global knowledge and requires a complex MIP solver to compute vehicle-to-platoon assignments. Overall, the distributed greedy approach achieves close to optimal results but requires the least assumptions and complexity. Therefore, we consider the distributed greedy approach the best approach among all presented approaches.