Direct Feature Evaluation in Black-Box Optimization Using Problem Transformations.

Area of Science:

• Optimization
• Machine Learning
• Data Science

Background:

• Exploratory Landscape Analysis (ELA) uses sample-based methods to characterize black-box optimization problems.
• Existing problem features aim to reveal landscape structures for effective algorithm selection.
• Current evaluation methods using machine learning introduce complexity and uncertainty.

Purpose of the Study:

• To propose a novel, direct method for evaluating the utility of problem features in optimization.
• To assess features without relying on additional machine learning models or confounding factors.
• To provide a quantitative measure of how well features detect problem landscape similarities.

Main Methods:

• The proposed method evaluates features based on their ability to detect a priori problem similarity rankings.
• Analysis of Variance (ANOVA) significance tests are employed to assess feature discrimination between successive problems.
• Features are assigned percentage scores for specific landscape characteristics based on ANOVA results.

Main Results:

• The novel evaluation method was demonstrated on twelve features across four problem transformations.
• Quantitative evidence was provided on the efficacy of different problem features in detecting landscape properties.
• The method successfully assigned scores indicating feature performance in identifying specific problem characteristics.

Conclusions:

• The proposed method offers a direct and less complex approach to evaluating optimization problem features.
• This technique facilitates a clearer understanding of feature utility for algorithm selection.
• The findings support the quantitative assessment of problem features for black-box optimization challenges.