MAGE.jl

MAGE is a type-aware extension of Cartesian Genetic Programming for evolving multimodal programs.

The Julia package that implements MAGE is still named UTCGP. Renaming the package is a separate migration because it affects the source tree, documentation, CI, and downstream users.

What MAGE Does

MAGE evolves typed computational graphs that can mix several modalities in the same program. A typical pipeline can take an image, apply a blur, then a Sobel operator, reduce the result to a scalar statistic, and use that scalar as a parameter for another image-processing function.

This typed composition is the core of the library:

• several data types can coexist in one evolved program
• function selection is constrained by input and output types
• image processing, scalar transforms, and control logic can be combined in the same graph

Current Scope

The current package already includes infrastructure for:

• typed graph construction and decoding
• FunctionBundle, Library, and MetaLibrary
• mutation operators and GA-based search
• search-network tooling and experiment tracing
• multimodal function libraries over images, floats, integers, strings, lists, tuples, and element operators

Existing image-oriented bundles in src/libraries/image2D/ cover:

• basic image operators
• filtering
• morphology
• binarization
• segmentation
• arithmetic and boolean arithmetic
• transcendental transforms
• experimental masking operators

There are also premade bundle collections for image-heavy setups, including Atari-oriented variants, in src/libraries/pre_made_libraries.jl.

Installation

Until the package rename happens, install and import the current package name:

using Pkg
Pkg.add(url="https://github.com/camilodlt/SearchNetworks.jl")
Pkg.add(url="https://github.com/camilodlt/MAGE.jl")
using UTCGP

For local development:

using Pkg
Pkg.develop(path=".")
Pkg.instantiate()

Publications

Papers defining MAGE:

• De La Torre, C., Lavinas, Y., Cortacero, K., Luga, H., Wilson, D. G., and Cussat-Blanc, S. "Multimodal Adaptive Graph Evolution." GECCO 2024. https://dl.acm.org/doi/abs/10.1145/3638530.3654347
• De La Torre, C., Lavinas, Y., Cortacero, K., Luga, H., Wilson, D. G., and Cussat-Blanc, S. "Multimodal Adaptive Graph Evolution for Program Synthesis." PPSN 2024. https://link.springer.com/chapter/10.1007/978-3-031-70055-2_19

Papers using MAGE for computer vision:

• De La Torre, C., Nadizar, G., Lavinas, Y., Schwob, R., Franchet, C., Luga, H., Wilson, D. G., and Cussat-Blanc, S. "Evolution of Inherently Interpretable Visual Control Policies." GECCO 2025. https://dl.acm.org/doi/abs/10.1145/3712256.3726332
• De La Torre, C., Nadizar, G., Lavinas, Y., Schwob, R., Franchet, C., Luga, H., Wilson, D. G., and Cussat-Blanc, S. "Evolved and Transparent Pipelines for Biomedical Image Classification." EuroGP 2025. https://link.springer.com/chapter/10.1007/978-3-031-89991-1_11