Data for the paper "Relating training instances to Automatic Design of Algorithms for bin packing via features"

Abstract

Automatic Design of Algorithms (ADA) shifts the burden of algorithm choice and design from developer to machine. Constructing an appropriate solver from a set of problem instances becomes a machine learning problem, with instances as training data. An efficient solver is trained for unseen problem instances with similar characteristics to those in the training set. However, this paper reveals that, as with classification and regression, for ADA not all training sets are equally valuable.

We apply a typical genetic programming ADA approach for bin packing problems to several new and existing public benchmark sets. Algorithms trained on some sets are general and apply well to most others, whereas some training sets result in highly specialised algorithms that do not generalise.

We relate these findings to features (simple metrics) of instances. Using instance sets with narrowly-distributed features for training results in highly specialised algorithms, whereas those with well-spread features result in very general algorithms. We show that variance in certain features has a strong correlation with the generality of the trained policies.

Our results provide further grounding for recent work using features to predict algorithm performance, and show the suitability of particular instance sets for training in ADA for bin packing.

This dataset includes the raw experimental data, new benchmark instances and figures for all experiments in the paper.