GeezerVison

OFL – Optimisation feature library

Submit features and corrections on GitHub with pull requests / issues, or by email: [email protected]

name	domain	objectives	variable type	constraints	reference	implementation	textual description
Exploratory Landscape Analysis (ELA)	black-box	1	continuous	unconstrained or box-constrained	https://doi.org/10.1145/2001576.2001690	https://github.com/Reiyan/pflacco	Exploratory Landscape Analysis (ELA) features
Cell Mapping Features	black-box	1	continuous	unconstrained or box-constrained	https://doi.org/10.1007/978-3-319-07494-8_9	https://github.com/Reiyan/pflacco	Cell Mapping Techniques for Exploratory Landscape Analysis
Dispersion Features	black-box	1	continuous	unconstrained or box-constrained	https://doi.org/10.1145/1143997.1144085	https://github.com/Reiyan/pflacco	The Dispersion Metric
Information Content-Based Features	black-box	1	continuous	unconstrained or box-constrained	https://doi.org/10.1109/TEVC.2014.2302006	https://github.com/Reiyan/pflacco	Exploratory Landscape Analysis of Continuous Space Optimization Problems Using Information Content. Information Content of Fitness Sequences (ICoFiS).
Nearest Better Features	black-box	1	continuous	unconstrained or box-constrained	https://doi.org/10.1145/2739480.2754642	https://github.com/Reiyan/pflacco	The Nearest-Better Features - also called Nearest-Better Clustering (NBC) Features. Features for detecting funnel structures.
Barrier Tree Features	black-box	1	continuous	unconstrained or box-constrained	https://doi.org/10.1007/978-3-030-25147-5_7	https://github.com/kerschke/flacco	Following generalized cell mapping, barrier trees can be used to characterize the problem structure.
Principal Components Features	black-box	1	continuous	unconstrained or box-constrained	https://doi.org/10.1007/978-3-030-25147-5_7	https://github.com/Reiyan/pflacco	Describe the variable scaling of continuous problems by applying a Principal Component Analysis.
Linear model features	black-box	1	continuous	unconstrained or box-constrained	https://doi.org/10.1007/978-3-030-25147-5_7	https://github.com/kerschke/flacco	Following generalized cell mapping, barrier trees can be used to characterize the problem structure.
Multi-objective Combinatorial Problem Features	black-box	2+	binary/combinatorial	unconstrained	https://doi.org/10.1162/evco_a_00193	?	Features describing Multi-objective Combinatorial Problems.
Local Optima Networks (LON)	black-box	1	combinatorial	unconstrained	https://doi.org/10.1007/978-3-642-41888-4_9	?	Local Optima Networks (LON): A network-based model of combinatorial landscapes.
DynamoRep	black-box	1	continuous	unconstrained	https://doi.org/10.1162/evco_a_00370	?	DynamoRep computes basic descriptive statistics such as min, max, mean, std over x and y values to capture algorithm-problem interaction in single-objective continuous optimization.
Opt2Vec	black-box	1	continuous	unconstrained	https://doi.org/10.1016/j.ins.2024.121134	?	A continuous optimization problem representation based on algorithm behavior.
DoE2Vec	black-box	1	continuous	unconstrained	https://doi.org/10.1145/3583133.3590609	https://github.com/nikivanstein/doe2vec	DoE2Vec: Deep-learning Based Features for Exploratory Landscape Analysis. Uses autoencoder (AE) based latent-space features. Does not require any feature engineering and is easily applicable to high-dimensional search spaces.
Best-so-far trajectories	black-box	1	continuous	unconstrained	https://doi.org/10.1007/978-3-031-56852-7_7	?	Use raw optimisation (probing) trajectories containing the best performance values seen so far for each evaluation.
Current trajectories	black-box	1	continuous	unconstrained	https://doi.org/10.1007/978-3-031-56852-7_7	?	Use raw optimisation (probing) trajectories containing the current performance values seen at each evaluation.
Time series features using TSFRESH	black-box	1	continuous	unconstrained	https://doi.org/10.1145/3449639.3459399	?	Compute time series features using TSFRESH. The authors did this for internal CMA-ES variables, but this might also be applied to, e.g., performance trajectories.
Deep learning-based features	black-box	1	continuous	unconstrained	https://doi.org/10.1145/3512290.3528834	?	Represents the initially sampled points as point clouds to 2D images for use with deep-learning models.
Deep-ELA	black-box	1, 2+	continuous	unconstrained	https://doi.org/10.1162/evco_a_00372	https://github.com/mvseiler/deep_ela	Deep-ELA: Deep Exploratory Landscape Analysis with Self-Supervised Pretrained Transformers for Single-Objective and Multiobjective Continuous Optimization Problems.
Decomposition-Based Multi-objective Landscape Features	black-box	2+	continuous	unconstrained	https://doi.org/10.1007/978-3-030-72904-2_3	?	A set of landscape features for multi-objective combinatorial optimization, by decomposing the original multi-objective problem into a set of single-objective sub-problems.
SOO search tree	black-box	1	continuous	unconstrained	https://doi.org/10.1145/3299904.3340308	?	Features on the basis of the search tree constructed by the so-called SOO global optimizer.
Scalarization-based features	black-box	2+	continuous	unconstrained	https://doi.org/10.1145/3712256.3726378	?	Scalarization-based ELA (S-ELA) for multi-objective continuous optimization using objective scalarization methods: decomposition and non-dominated sorting.
Landscape Features for Multi-objective Interpolated Continuous Optimisation Problems	black-box	2+	continuous	unconstrained	https://doi.org/10.1145/3449639.3459353	?	Landscape Features for Multi-objective Interpolated Continuous Optimisation Problems aimed at measuring aspects of: global properties, multimodality, evolvability, and ruggedness. Uses a different sampling method to use features originally designed for discrete problems for the continuous case.
Fitness distance correlation (FDC)	white-box, grey-box	1	continuous, combinatorial	unconstrained	Jones and Forrest 1995 - Fitness Distance Correlation as a Measure of Problem Difficulty for Genetic Algorithms	?	Correlation between performance values and the distance to the global optimum. I.e., the global optimum must be known to compute this feature.
Evolvability metrics	black-box	1	continuous, combinatorial	unconstrained	https://doi.org/10.1162/106365602317301754	?	Evolvability statistics of the fitness landscape surrounding sampled solutionas, to assess the ruggedness and neutrality of a landscape.
Bag of local landscape patterns (BoLLP)	black-box	1	continuous, combinatorial	unconstrained	https://doi.org/10.1007/s00500-016-2091-4	?	Uses the fitness patterns of local candidate solutions.
Bag of evolvability (BoEvo)	black-box	1	continuous, combinatorial	unconstrained	https://doi.org/10.1007/s00500-016-2091-4	?	Uses the evolvability of candidate solutions as local landscape feature.
name	domain	objectives	variable type	constraints	reference	implementation	textual description