Instance Space Analysis for the Generalized Assignment Problem

Tobias Geibinger, Lucas Kletzander, Nysret Musliu

Abstract: In this work, we consider the well-studied Generalized Assignment Problem and investigate the performance of several metaheuristic methods. To obtain insights on strengths and weaknesses of these solution approaches, we perform Instance Space Analysis on the existing instance types and propose a set of features describing the hardness of an instance. This is of interest since the existing benchmark set is dated and rather limited and the known instance generators might not be fully representative. Our analysis for metaheuristic methods reveals that this is indeed the case and finds several gaps, which we fill with newly generated instances thus adding diversity and providing a new benchmark instance set. Furthermore, we analyze the impact of problem features on the performance of the methods used and identify the most important ones.