Improvements for mlrose applied to the Traveling Salesperson Problem

TL;DR

The paper tackles optimizing an industrial order-picking task, framed as a Traveling Salesperson Problem, by applying and improving the mlrose library's GA and HC implementations. It introduces a reversal-invariant crossover for GA to respect the symmetry of tours and a down-step plus visited-state mechanism for HC to mitigate local optima, alongside fixing a negative-fitness bug in mlrose. Experimental results on the att48 TSPLIB dataset show substantial GA gains (roughly 46% and 39% reductions in tour length) and modest HC gains (approximately 2.1% and 4.6%), with computation times increasing accordingly. The work emphasizes that leveraging problem structure is essential when repurposing generic optimization libraries for real-world tasks, and that understanding the underlying methods remains important even as automation aids hyperparameter tuning.

Abstract

In this paper we discuss the application of Artificial Intelligence (AI) to the exemplary industrial use case of the two-dimensional commissioning problem in a high-bay storage, which essentially can be phrased as an instance of Traveling Salesperson Problem (TSP). We investigate the mlrose library that provides an TSP optimizer based on various heuristic optimization techniques. Our focus is on two methods, namely Genetic Algorithm (GA) and Hill Climbing (HC), which are provided by mlrose. We present improvements for both methods that yield shorter tour lengths, by moderately exploiting the problem structure of TSP. That is, the proposed improvements have a generic character and are not limited to TSP only.

Figures (1)

• Figure 1: Results of the experiments on the (Fig. \ref{['fig:gen_distance_plot']}) and (Fig. \ref{['fig:hill_distance_plot']}).