Learning Heuristic Selection with Dynamic Algorithm Configuration

Speck, D. R.; Biedenkapp, André; Hutter, Frank; Mattmüller, Robert; Lindauer, Marius

doi:10.1609/icaps.v31i1.16008

Cited by 10 publications

(18 citation statements)

References 25 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These application domains, together with the learning rate control setting from (Daniel et al, 2016), have later been released as part of a benchmark suite, called DACbench (Eimer et al, 2021b), offering a unified interface that facilitates comparisons between different DAC methods across different DAC scenarios. In this article, we extend this initial discussion of Biedenkapp et al (2020) and present a thorough empirical comparison of AC and DAC on these three different realworld DAC applications (Daniel et al, 2016;Shala et al, 2020;Speck et al, 2021) using the unified DACbench interface.…”

Section: Prior Art: Automated Dynamic Algorithm Configurationmentioning

confidence: 67%

“…DAC has been an active research area that has produced various highly practical algorithms leveraging dynamic parameter adaptation mechanisms to empirically outperform their static counterparts, e.g., Reactive Tabu Search (Battiti & Tecchiolli, 1994), CMA-ES (Hansen et al, 2003), and Adam (Kingma & Ba, 2015). Beyond these empirical successes and dedicated case studies (e.g., Senior et al, 2013;, the potential of DAC has also been shown theoretically (Moulines & Bach, 2011;Warwicker, 2019;Speck et al, 2021).…”

Section: Introductionmentioning

confidence: 90%

“…3. applied DAC to AI planning, tackling the problem of heuristic selection in FastDownward (Helmert, 2006), and showed how DAC subsumes static algorithm configuration and can improve upon the best possible algorithm selector (Speck et al, 2021); and 4. presented DACBench, the first benchmark library for DAC, facilitating reproducible results through a unified interface (Eimer et al, 2021b).…”

Section: Summary Of Contributionsmentioning

confidence: 99%

“…Biedenkapp et al (2020) introduced DAC in an attempt to consolidate these isolated efforts and to raise the level of generality in pursuit of algorithms similar to those that exist for static AC. Direct follow-up work has provided additional evidence for the practicality of DAC by learning step-size adaptation in CMA-ES (Shala et al, 2020), and by learning to select heuristics in the FastDownward planner (Speck et al, 2021). These application domains, together with the learning rate control setting from (Daniel et al, 2016), have later been released as part of a benchmark suite, called DACbench (Eimer et al, 2021b), offering a unified interface that facilitates comparisons between different DAC methods across different DAC scenarios.…”

Section: Prior Art: Automated Dynamic Algorithm Configurationmentioning

confidence: 99%

“…Modern RL Over the last decade, a series of methodological advances have given rise to a new generation of RL methods that can tackle complex real-world problems (Mnih Silver et al, 2016;Barozet et al, 2020;Lee et al, 2020), and that have also been successfully applied to DAC. In particular, modern RL methods based on deep neural networks can effectively learn useful representations that allow them to handle complex state and action spaces, using, e.g., (double) deep Q-learning (DDQN, Hansen, 2016;Sharma et al, 2019;Speck et al, 2021;Bhatia et al, 2021), modern actor critic (Ichnowski et al, 2021), and policy gradient methods (Daniel et al, 2016;Xu et al, 2019;Gomoluch et al, 2019;Shala et al, 2020;Getzelman & Balaprakash, 2021;Almeida et al, 2021).…”

Section: Dac By Reinforcement Learningmentioning

confidence: 99%

See 4 more Smart Citations

Automated Dynamic Algorithm Configuration

Adriaensen

Biedenkapp

Shala

et al. 2022

jair

View full text Add to dashboard Cite

The performance of an algorithm often critically depends on its parameter configuration. While a variety of automated algorithm configuration methods have been proposed to relieve users from the tedious and error-prone task of manually tuning parameters, there is still a lot of untapped potential as the learned configuration is static, i.e., parameter settings remain fixed throughout the run. However, it has been shown that some algorithm parameters are best adjusted dynamically during execution. Thus far, this is most commonly achieved through hand-crafted heuristics. A promising recent alternative is to automatically learn such dynamic parameter adaptation policies from data. In this article, we give the first comprehensive account of this new field of automated dynamic algorithm configuration (DAC), present a series of recent advances, and provide a solid foundation for future research in this field. Specifically, we (i) situate DAC in the broader historical context of AI research; (ii) formalize DAC as a computational problem; (iii) identify the methods used in prior art to tackle this problem; and (iv) conduct empirical case studies for using DAC in evolutionary optimization, AI planning, and machine learning.

show abstract

Section: Prior Art: Automated Dynamic Algorithm Configurationmentioning

confidence: 67%

Section: Introductionmentioning

confidence: 90%

Section: Summary Of Contributionsmentioning

confidence: 99%

Section: Prior Art: Automated Dynamic Algorithm Configurationmentioning

confidence: 99%

Section: Dac By Reinforcement Learningmentioning

confidence: 99%

See 3 more Smart Citations

Automated Dynamic Algorithm Configuration

Adriaensen

Biedenkapp

Shala

et al. 2022

jair

View full text Add to dashboard Cite

show abstract

Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges

Bischl

Binder

Lang

et al. 2023

WIREs Data Min & Knowl

Self Cite

216

View full text Add to dashboard Cite

Most machine learning algorithms are configured by a set of hyperparameters whose values must be carefully chosen and which often considerably impact performance. To avoid a time-consuming and irreproducible manual process of trial-anderror to find well-performing hyperparameter configurations, various automatic hyperparameter optimization (HPO) methods-for example, based on resampling error estimation for supervised machine learning-can be employed. After introducing HPO from a general perspective, this paper reviews important HPO methods, from simple techniques such as grid or random search to more advanced methods like evolution strategies, Bayesian optimization, Hyperband, and racing. This work gives practical recommendations regarding important choices to be made when conducting HPO, including the HPO algorithms themselves, performance evaluation, how to combine HPO with machine learning pipelines, runtime improvements, and parallelization.

show abstract

Realtime gray-box algorithm configuration using cost-sensitive classification

Weiss,

Tierney

2023

Ann Math Artif Intell

View full text Add to dashboard Cite

A solver’s runtime and the quality of the solutions it generates are strongly influenced by its parameter settings. Finding good parameter configurations is a formidable challenge, even for fixed problem instance distributions. However, when the instance distribution can change over time, a once effective configuration may no longer provide adequate performance. Realtime algorithm configuration (RAC) offers assistance in finding high-quality configurations for such distributions by automatically adjusting the configurations it recommends based on instances seen so far. Existing RAC methods treat the solver as a black box, meaning the solver is given a configuration as input, and it outputs either a solution or runtime as an objective function for the configurator. However, analyzing intermediate output from the solver can enable configurators to avoid wasting time on poorly performing configurations. We propose a gray-box approach that utilizes intermediate output during evaluation and implement it within the RAC method Contextual Preselection with Plackett-Luce (CPPL blue). We apply cost-sensitive machine learning with pairwise comparisons to determine whether ongoing evaluations can be terminated to free resources. We compare our approach to a black-box equivalent on several experimental settings and show that our approach reduces the total solving time in several scenarios and improves solution quality in an additional scenario.

show abstract

Learning Heuristic Selection with Dynamic Algorithm Configuration

Cited by 10 publications

References 25 publications

Automated Dynamic Algorithm Configuration

Automated Dynamic Algorithm Configuration

Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges

Realtime gray-box algorithm configuration using cost-sensitive classification

Contact Info

Product

Resources

About