A Machine Learning Approach for Modeling Algorithm Performance Predictors

Abstract

This paper deals with heuristic algorithm selection, which can be stated as follows: given a set of solved instances of a NP-hard problem, for a new instance to predict which algorithm solves it better. For this problem, there are two main selection approaches. The first one consists of developing functions to relate performance to problem size. In the second more characteristics are incorporated, however they are not defined formally, neither systematically. In contrast, we propose a methodology to model algorithm performance predictors that incorporate critical characteristics. The relationship among performance and characteristics is learned from historical data using machine learning techniques. To validate our approach we carried out experiments using an extensive test set. In particular, for the classical bin packing problem, we developed predictors that incorporate the interrelation among five critical characteristics and the performance of seven heuristic algorithms. We obtained an accuracy of 81% in the selection of the best algorithm.

Small Boxes Big Data: A Deep Learning Approach to Optimize Variable Sized Bin Packing

“…J. Perez O. et al [7] propose a methodology to model the algorithm performance predictor. Given a set of solved instances of a NP-hard problem and set of heuristic algorithms, the authors model the relationship among performance and characteristics to predict which algorithm solves the problems better.…”

“…We incorporate prior knowledge to speed up the learning procedure and increase the prediction accuracy by the predefined features. Instead of applying dimension reduction processing like PCA [6], Factor Analysis [7] and NMF [19], we use all the raw features listed in Table III. The first seven features are from the items and the next six features are from the bins.…”

“…Our network instance is defined by (8) to (12) based on equation (5) to (7). m i represents the number of nodes at level l. We apply softplus as the activation function to all the hidden nodes except for the last layer where we apply softmax, as it is designed to model a classification predictor.…”

“…In most cases, it is a trade off between the solution quality and the time to solve it. Adapting to different problem instances by different algorithms has been studied by traditional AI machine learning approaches [6] [7] [8] [9]. These works initially optimize the packing algorithm end-to-end and later to a finer granularity, state-to-state.…”

A machine learning approach for automated strip packing algorithm selection

“…The use of data mining approaches in cutting and packing problems are mostly related to the process of converting problem information into measurable factors in order to reflect the main problem characteristics and compare algorithm performance with different types of problem instances [32]. In a rare attempt at predicting bin packing algorithm performance predictors, Perez et al [38] proposed a methodology that model the relation-ship between algorithm performance and characteristics of bin packing problem instances using machine learning techniques. In [47], Smith-Miles & Lopes proposed a methodology for adequately characterising the features of a problem instance and showed how such features can be defined and measured for various optimisation problems including the bin packing problems.…”