🤖 machine learning

Beyond Numerical Features: CNN-Driven Algorithm Selection via Contour Plots for Continuous Black-Box Optimization

This paper proposes a novel per-instance algorithm selection method for continuous black-box optimization that utilizes a CNN to analyze contour plot visualizations of probed landscapes, demonstrating that this image-based approach outperforms the single best solver and remains competitive with traditional feature-based methods without relying on handcrafted descriptors.

Original authors: Yiliang Yuan, Xiang Shi, Mustafa Misir

Published 2026-05-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yiliang Yuan, Xiang Shi, Mustafa Misir

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef trying to bake the perfect cake. You have a huge pantry full of different recipes (algorithms), but you don't know which one works best for the specific ingredients you have in front of you today. Some recipes work great with flour and eggs, while others are better for chocolate and nuts.

In the world of computer science, this is called Continuous Black-Box Optimization. You have a "black box" (a complex problem) where you can only taste the result (get a score) but can't see the recipe inside. The goal is to pick the best "solver" (recipe) for the specific problem you are facing.

The Old Way: Reading a List of Numbers

Traditionally, computers tried to solve this by taking a few samples of the problem and turning them into a long list of numbers (like "it's bumpy," "it's curved," "it's spiky"). These are called numerical features. It's like trying to describe a mountain range by only reading a list of its average height, slope, and temperature. It gives you data, but it misses the big picture.

The New Way: Looking at a Map

This paper proposes a simpler, more visual approach. Instead of turning the problem into a list of numbers, the authors turn it into a picture.

Think of the problem as a hilly landscape. The authors take a "probe" (a set of measurements) and draw a contour map of that landscape, just like a hiking map showing the peaks and valleys.

The Input: They generate these maps using a grid of 300x300 points.
The Brain: They feed these pictures into a CNN (Convolutional Neural Network). You can think of a CNN as a super-smart robot that is really good at looking at pictures and spotting patterns, just like how your brain recognizes a face in a crowd.

How It Works

The Setup: They have a portfolio of 12 different "solver" algorithms (the 12 recipes).
The View: For every new problem, they generate a few different "views" (contour maps) of the landscape.
- For 2D problems: They see the whole map.
- For complex 3D+ problems: They take a "slice" of the high-dimensional space to create a 2D picture, like slicing a loaf of bread to see the texture inside.
The Prediction: The CNN looks at these pictures and predicts: "If I use Recipe A, I'll get a score of X. If I use Recipe B, I'll get a score of Y."
The Choice: The system picks the recipe predicted to give the best score.

What They Found

The researchers tested this on a standard set of difficult math problems (called BBOB).

Beating the "One-Size-Fits-All": They compared their visual system against the "Single Best Solver" (SBS)—which is just picking the one best recipe that works on average for everything. Their visual system crushed the SBS, finding the right tool for the specific job much more often.
Competing with the Experts: They also compared it to the old "list of numbers" methods (ELA and Deep-ELA). Their picture-based method performed just as well, and sometimes even better, especially on medium-difficulty problems.
Resolution Matters: They found that looking at a higher-resolution picture (300x300 pixels) helped the robot make better choices than a blurry, low-resolution one (64x64 pixels), though it took a bit more computer power to process.

The Limits (The "Fine Print")

The authors are honest about where this method has limits:

It's a bit expensive to make the map: Generating these high-quality pictures requires a lot of initial "tasting" (calculations). They admit this is great for offline planning (where you have time to prepare) but might be too slow for real-time, split-second decisions.
The "Slice" problem: For very complex, high-dimensional problems, a single 2D slice of the map might miss some hidden details, which is why it didn't win on the absolute hardest problems.
Specific to this test: They tested this on a specific set of problems and a specific list of 12 solvers. It's a proof that "pictures work," but it hasn't been tested on every possible type of problem in the world yet.

The Bottom Line

This paper shows that you don't always need to turn a complex problem into a boring list of numbers to solve it. Sometimes, just showing the problem to a computer as a picture allows it to "see" the structure of the landscape and pick the perfect tool for the job, often outperforming the old, number-heavy methods.

Technical Summary: CNN-Driven Algorithm Selection via Contour Plots

Problem Statement
Continuous black-box optimization (BBO) involves minimizing or maximizing objective functions where gradient information is unavailable, relying solely on function evaluations. While derivative-free solvers (e.g., CMA-ES variants) perform well, the No Free Lunch theorem dictates that no single algorithm universally outperforms others across all problem instances. Consequently, Automated Algorithm Selection (AAS) aims to choose the most suitable solver from a fixed portfolio for a specific problem instance.

Existing AAS approaches in continuous BBO predominantly rely on Exploratory Landscape Analysis (ELA), which constructs numerical feature vectors (e.g., curvature, separability) from probed evaluations, or Deep-ELA, which uses learned embeddings from pretrained transformers. The authors argue that these numerical descriptors may fail to capture the spatial structure evident in visualized landscapes. While CNN-based AAS has been explored in discrete domains with direct instance encodings, its application to continuous BBO—where representations must be constructed via probing—is under-explored.

Methodology
The paper proposes a representation-driven AAS approach that treats probed landscapes as images rather than numerical vectors.

Instance Representation (Contour Plots):
- Single-Objective (SOO): Each problem instance is rendered as a 2D grayscale contour map by evaluating the objective function on a fixed $300 \times 300$ grid. For dimensions $d > 2$ , a 2D slice is created by randomly selecting two coordinates to span a subspace while fixing others to zero (leveraging BBOB's symmetric domain).
- Multi-Objective (MOO): For bi-objective problems ( $d=2, m=2$ ), a sub-contour sampling strategy is used. Five axis-aligned rectangular windows are sampled within the domain for each objective, generating multiple views to reduce reliance on a single global rendering.
- Input Variants: To study input fidelity, plots are resized to resolutions $r \in \{64, 128, 300\}$ .
CNN Architectures:
Two model variants are proposed to predict the relative performance (Expected Running Time for SOO, relative Hypervolume for MOO) of algorithms in a portfolio:
- Combined Model: The five instance-specific views (or window samples) are stacked along the channel dimension and fed directly into a CNN encoder.
- Separate Model: Each view is processed individually by a shared-weight encoder, and the resulting feature vectors are concatenated before being passed to a regression head.
- SOO Architecture: A three-layer CNN encoder.
- MOO Architecture: A ResNet-18 encoder due to higher representational demands.
Selection Mechanism:
The model predicts the performance metric (e.g., relative ERT) for each algorithm in the portfolio. The solver with the best predicted value (minimum ERT or maximum HV) is selected.

Key Contributions

Probing-Based Image Representation: The authors introduce a novel AAS formulation for continuous BBO that uses contour-map renderings of probed landscapes as instance representations, eliminating the need for handcrafted ELA features.
Empirical Evaluation: The study provides a comprehensive analysis on standard BBOB 2009 single-objective benchmarks and a bi-objective setting following the Deep-ELA protocol. It analyzes the effects of view aggregation (Combined vs. Separate) and input resolution.
Performance Benchmarking: The work positions contour-based selection against established feature-based baselines (ELA with MLP/RF and Deep-ELA variants).

Experimental Results

Single-Objective Optimization (SOO):
- On the BBOB 2009 benchmark (96 configurations), the Combined CNN with $300 \times 300$ input resolution significantly outperformed the Single Best Solver (SBS), reducing the mean relative ERT from 30.37 (SBS) to 5.60.
- The CNN approach was competitive with the strongest feature-based baselines (ELA-MLP: 5.72; Deep-ELA Medium-kNN: 6.04).
- Performance generally improved with higher input resolution ( $r=300$ ), though this increased training costs.
- The method showed strong gains on function groups F1–F19, while performance on the most difficult group (F20–F24) remained mixed compared to feature-based methods.
Multi-Objective Optimization (MOO):
- Under the Deep-ELA evaluation setting, CNN-based selectors were competitive with Deep-ELA baselines.
- The Separate CNN achieved the best mean relative HV scores (0.971–0.974), outperforming Deep-ELA variants (0.904).
- The method was particularly effective on ZDT instances (achieving near-perfect scores), while results on DTLZ and MMF were strong but slightly below the best Deep-ELA settings.

Significance and Claims
The paper claims that simple vision models can exploit the spatial structure of probed landscapes for algorithm selection without relying on handcrafted ELA features. The results suggest that contour-plot visualizations contain sufficient signal to drive effective AAS in continuous BBO, offering a complementary representation to numerical descriptors.

Limitations and Scope
The authors explicitly note several limitations:

Probing Budget: The current approach requires a substantial probing budget (five $300 \times 300$ maps per configuration), making it suitable for offline analysis but not strict low-budget online deployment.
High-Dimensional Representation: For $d > 2$ , the 2D slice representation is a partial view of the landscape, which may contribute to weaker performance on the most difficult function groups.
Fixed Protocols: The evaluation is tied to specific portfolios and protocols (including window sampling for MOO), and the MOO study is limited to $d=2, m=2$ .

The work is presented as a proof-of-concept demonstrating the feasibility of image-based representations, with future directions pointing toward cost-sensitive probing, multi-view designs, and hybrid visual-numerical features.