Evolver-Studio: a graphical interface for Evolver

Evolver-Studio is a Python/Streamlit application for Evolver, the Java framework for the automated meta-optimization of multi-objective metaheuristics. Evolver has no GUI: its algorithms are configured through YAML parameter spaces, and training runs are launched from Java code or its command-line entry point. Evolver-Studio makes Evolver usable without writing Java or editing YAML by hand, and adds an analysis layer on top of its results.

It serves two purposes:

• Solving problems: configure and run Evolver's configurable algorithms on a concrete problem, in the style of jMetal's runners. Evolver's configurable core can be used on its own, as an alternative to jMetal.
• Meta-optimization: launch and monitor training runs, then analyze and validate the configurations they find.

Status: early development, heading to 0.1.0 (see CHANGELOG.md). Exploring parameter spaces, launching and monitoring training runs, analyzing them, running an algorithm on a problem, validating a tuned configuration, and five tutorials are available. See ROADMAP.md.

Evolver-Studio runs Evolver's release jar, downloaded from Maven Central, as a subprocess, and talks to it through files rather than a server:

Evolver-Studio (Streamlit)
  └─> writes request.yaml (+ reusable baseLevel / metaSearch files)
       └─> java -cp Evolver-2.4-jar-with-dependencies.jar
                org.uma.evolver.cli.training.TrainingRunnerMain request.yaml status.yaml
            ├─> status.yaml   polled for progress (RUNNING / FINISHED / FAILED)
            └─> METADATA.txt, INDICATORS.csv, CONFIGURATIONS.csv, VAR_CONF.txt

Training runs are long batch jobs, so each one runs as a detached process: closing or reloading the browser does not stop it, and reopening the app reconnects to it. The parameter spaces the app shows and edits are read from the jar itself, and the reference fronts and weight vectors that training runs need are copied from Evolver into resources/. The catalogue of algorithms, problems and indicators is cross-checked against the manifest printed by Evolver's org.uma.evolver.cli.training.DescribeMain.

The app opens on a home page with a card for each part of the tool. The menu groups the pages by purpose:

✅ available · 🚧 planned

The Training page provides:

• A parameter editor for the base-level algorithm, with a guided mode (forms, valid by construction) and an expert mode (raw YAML, validated on every change).
• Training sets of one or more problems, each with its reference front and evaluation budget. The problems are chosen from a list (those of the base algorithm's encoding), or from the same listing as Explore › Problems, with each problem's dimensions, arguments and fronts in view; any other jMetal problem can be added by class name. A problem can be given its constructor's arguments (e.g. DTLZ2 with 2 objectives), and one of another encoding than the base algorithm's is reported before launching.
• A choice of meta-optimizer, with its operator settings pre-filled from Evolver's example files.
• A monitor for runs that last hours: progress, pace and time left (with a warning when a run seems stuck), the front and, optionally, the meta-optimizer's whole population as it evolves (a checkbox before launching), how each meta-objective converges, the configurations found so far (downloadable, to validate one without waiting for the end) and the runner's log. It follows the files Evolver appends to without reading them again at every refresh, and survives closing the page: the run is picked up again when it is reopened.

Evolver-Studio can browse the parameter space of every configurable algorithm in Evolver. The ones that can be launched from the app today are:

MOPSO is browsable only, until Evolver's command-line runner supports it. See Evolver's supported algorithms for the full list.

• Python 3.11+ and Conda
• Java 21 or newer, on the PATH

Neither an Evolver checkout nor Maven is needed: the app downloads the jar of Evolver 2.4 (the release the app is built against) from Maven Central. Other versions are not supported; the expected version is set by EVOLVER_VERSION in evolver_studio/evolver_client.py.

Evolver-Studio is versioned on its own, and works with one Evolver release at a time:

git clone https://github.com/jMetal/Evolver-Studio.git
cd Evolver-Studio
make env   # creates (or updates) the 'evolver-studio' environment from environment.yml

make run   # streamlit run app.py

1. The first time, click Download Evolver 2.4 in the sidebar. The jar (about 130 MB) is downloaded from Maven Central into lib/ and its checksum is verified.
2. The home page shows the parts of the app. Open Base algorithms, under Explore, to browse the algorithms and their parameter spaces.
3. Open Training, keep the default settings (NSGA-II on ZDT4) or edit them, and click Launch training. The indicator front updates as the run progresses.
4. Paths are relative to the Evolver-Studio directory. Each run's request files (request.yaml, status.yaml, ...) are kept under cli-runner-runs/<run_id>/, and its results under <output directory>/<run_id>/ (by default, results/nsgaii/ZDT4/<run_id>/).

New to Evolver? Start with the Tutorials page.

make lint     # ruff check .
make format   # ruff format .
make test     # pytest tests/ -x

To run the app against another Evolver build, such as one made from Evolver's develop branch while working on Evolver itself, point the EVOLVER_JAR environment variable at its jar:

EVOLVER_JAR=/path/to/Evolver-2.5-SNAPSHOT-jar-with-dependencies.jar make run

Tests that use Evolver's jar (downloaded, or set by EVOLVER_JAR) are skipped without it. Among them, tests/test_catalogue.py keeps Evolver-Studio's hand-maintained catalogue (evolver_studio/catalogue.py) in sync with Evolver: it fails on any parameter space in the jar that the catalogue does not account for, and cross-checks the catalogue against the manifest printed by Evolver's DescribeMain. Evolver has matching tests for its algorithm and meta-optimizer registries.

Moving to a new Evolver release means changing EVOLVER_VERSION in evolver_studio/evolver_client.py and refreshing the copied resources with make sync-resources, which downloads them from the release's source archive on GitHub. tests/test_resources.py checks the copy against the checksums that command records.

Code and commits follow CODING_GUIDELINES.md and GIT_GUIDELINES.md.

Next steps include the analysis layer, and more statistics for validation studies (Friedman and critical difference plots). See ROADMAP.md for details and the use cases behind them.

Evolver-Studio has no publication of its own yet. If you use it in your research, please cite Evolver:

@article{AND23,
  title   = {Evolver: Meta-optimizing multi-objective metaheuristics},
  journal = {SoftwareX},
  volume  = {23},
  pages   = {101551},
  year    = {2024},
  issn    = {2352-7110},
  doi     = {10.1016/j.softx.2023.101551},
}

This project is licensed under the GNU General Public License v3.0 — see the LICENSE file for details.