# A Collection of Deep Learning-based Feature-Free Approaches for Characterizing Single-Objective Continuous Fitness Landscapes

**Authors:** Moritz Vinzent Seiler, Raphael Patrick Prager, Pascal Kerschke, Heike Trautmann  
**arXiv:** [2204.05752](https://arxiv.org/abs/2204.05752)  
**Date:** 2022-04-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2204.05752v2/

## Key points

- The paper discusses the use of Exploratory Landscape Analysis (ELA) for characterizing landscapes in single-objective continuous optimization problems.
- The authors propose a collection of different approaches to characterize optimization landscapes, ranging from point clouds to 2D images, that are entirely feature-free.
- Deep Learning techniques are utilized to predict high-level expert-based landscape properties such as multimodality degree and funnel structures.
- The quality of the proposed approaches is comparable to methods relying on traditional landscape features.
- This work provides an alternative perspective for various research areas involving optimization problems by eliminating the reliance on traditional landscape features.

## Summary

The paper titled "A Collection of Deep Learning-based Feature-Free Approaches for Characterizing Single-Objective Continuous Fitness Landscapes" discusses the use of Exploratory Landscape Analysis (ELA) as a technique for characterizing landscapes in single-objective continuous optimization problems. The authors highlight the importance of landscape insights for problem understanding and benchmark set assessment. To address the limitations and drawbacks of conventional landscape features, they propose a collection of different approaches to characterize optimization landscapes. These representations range from point clouds to 2D images and are entirely feature-free. The authors demonstrate and validate their methods on the BBOB testbed and utilize Deep Learning techniques to predict high-level expert-based landscape properties such as multimodality degree and funnel structures. The quality of their approaches is found to be comparable to methods relying on traditional landscape features. This provides an exciting new perspective for various research areas that utilize problem information, including problem understanding, algorithm design, automated algorithm configuration, and selection. In terms of methodology, the authors define a training dataset consisting of input data (XEla) and class labels (Y). The input data comprises landscape features while the class labels represent the high-level properties being considered. The experiments involve 62 landscape features and 14,400 observations generated using Latin Hypercube Sampling on a specific problem instance with missing values observed in 14 features belonging to ela_level and dispers feature classes. Overall, this paper presents innovative approaches that eliminate the reliance on traditional landscape features for characterizing optimization landscapes. The results obtained through Deep Learning techniques show promise in accurately predicting high-level properties without explicitly computing specific features. This work opens up new possibilities for various research areas involving optimization problems by providing alternative perspectives based on alternative representations of samples rather than traditional feature computation.