Algorithm selection using deep learning without feature extraction

Alissa, Mohamad; Sim, Kevin; Hart, Emma

doi:10.1145/3321707.3321845

Cited by 26 publications

(34 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As mentioned in the introduction, in a recent conference paper, Seiler et al (2020) adopted and adapted the LSTM approach we proposed in (Alissa et al, 2019) to be applicable to the Euclidean TSP domain, also using an evolved (and balanced) dataset (1000 instances) with two TSP solvers. They compared a feature-based approach using four different classical ML classifiers to a feature-free approach using deep learning Convolutional Neural Networks (CNNs).…”

Section: Asp With Deep Learning Approachesmentioning

confidence: 99%

“…In a recent conference paper (Alissa et al, 2019), we described an initial implementation of an RNN-LSTM to perform algorithm-selection in the field of 1D bin-packing, showing that it was able to outperform the Single-Best Solver (SBS) (i.e. the single heuristic that achieves the best performance over the instance set) on multiple datasets and achieving comparable performance to the Virtual-Best Solver (VBS), i.e.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Automated Algorithm Selection: from Feature-Based to Feature-Free Approaches

Alissa¹,

Sim²,

Hart³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

We propose a novel technique for algorithm-selection, applicable to optimisation domains in which there is implicit sequential information encapsulated in the data, e.g., in online bin-packing. Specifically we train two types of recurrent neural networks to predict a packing heuristic in online binpacking, selecting from four well-known heuristics. As input, the RNN methods only use the sequence of item-sizes. This contrasts to typical approaches to algorithm-selection which require a model to be trained using domain-specific instance features that need to be first derived from the input data. The RNN approaches are shown to be capable of achieving within 5% of the oracle performance on between 80.88% to 97.63% of the instances, depending on the dataset. They are also shown to outperform classical machine learning models trained using derived features. Finally, we hypothesise that the proposed methods perform well when the instances exhibit some implicit structure that results in discriminatory performance with respect to a set of heuristics. We test this hypothesis by generating fourteen new datasets with increasing levels of structure, and show that there is a critical threshold of structure required before algorithm-selection delivers benefit.

show abstract

Section: Asp With Deep Learning Approachesmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Automated Algorithm Selection: from Feature-Based to Feature-Free Approaches

Alissa¹,

Sim²,

Hart³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…When the item doesn't fit in any container, then the container with the lowest free space is closed and a new one is opened to pack/stack the item into." We use 900 bin-packing instances from datasets 3 first defined in [30], each of which has 120 items and is initialised with item sizes drawn from two different distributions. In order to solve this as a streaming instance, we consider the items to arrive in the order defined in each instance.…”

Section: Streaming Bin-packing Data Instancesmentioning

confidence: 99%

“…Each instance is solved best by one of the heuristics under investigation (according to the Falkenauer fitness function [31] given in equation 1). The heuristics considered are -Best-Fit; First-Fit and Worst-Fit [30] (BF, FF and WF). Thus, 150 instances are solved best by each one of them.…”

Section: Streaming Bin-packing Data Instancesmentioning

confidence: 99%

A Neural Approach to Generation of Constructive Heuristics

Alissa

Sim

Hart

2021

2021 IEEE Congress on Evolutionary Computation (CEC)

Self Cite

View full text Add to dashboard Cite

Both algorithm-selection methods and hyperheuristic methods rely on a pool of complementary heuristics. Improving the pool with new heuristics can improve performance, however, designing new heuristics can be challenging. Methods such as genetic programming have proved successful in automating this process in the past. Typically, these make use of problem state-information and existing heuristics as components. Here we propose a novel neural approach for generating constructive heuristics, in which a neural network acts as a heuristic by generating decisions. We evaluate two architectures, an Encoder-Decoder LSTM and a Feed-Forward Neural Network. Both are trained using the decisions output from existing heuristics on a large set of instances. We consider streaming instances of binpacking problems in a continual stream that must be packed immediately in strict order and using a limited number of resources. We show that the new heuristics generated are capable of solving a subset of instances better than the well-known heuristics forming the original pool, and hence the overall value of the pool is improved w.r.t. both Falkenauer's performance metric and the number of bins used.

show abstract

“…Generation of discriminating instances is increasingly popular [4], for instance to measure algorithm performance across instances [28] or to improve algorithm selection tools [30,14]. Most studies in this area tackle one problem class at a time.…”

Section: Related Workmentioning

confidence: 99%

Discriminating Instance Generation from Abstract Specifications: A Case Study with CP and MIP

Akgün

Dang

Miguel

et al. 2020

Integration of Constraint Programming, Artificial Intelligence, and Operations Research

View full text Add to dashboard Cite

We extend automatic instance generation methods to allow cross-paradigm comparisons. We demonstrate that it is possible to completely automate the search for benchmark instances that help to discriminate between solvers. Our system starts from a high level humanprovided problem specification, which is translated into a specification for valid instances. We use the automated algorithm configuration tool irace to search for instances, which are translated into inputs for both MIP and CP solvers by means of the Conjure, Savile Row, and MiniZinc tools. These instances are then solved by CPLEX and Chuffed, respectively. We constrain our search for instances by requiring them to exhibit a significant advantage for MIP over CP, or vice versa. Experimental results on four optimisation problem classes demonstrate the effectiveness of our method in identifying instances that highlight differences in performance of the two solvers.

show abstract

Algorithm selection using deep learning without feature extraction

Cited by 26 publications

References 22 publications

Automated Algorithm Selection: from Feature-Based to Feature-Free Approaches

Automated Algorithm Selection: from Feature-Based to Feature-Free Approaches

A Neural Approach to Generation of Constructive Heuristics

Discriminating Instance Generation from Abstract Specifications: A Case Study with CP and MIP

Contact Info

Product

Resources

About