MO-CMA-MAE

This repository implements the multi-objective quality-diversity algorithm MO-CMA-MAE and baselines over a set of benchmark problems.

Requirements

• Python 3.11+
• uv for dependency management

Installation

From the repository root:

uv sync

Running an experiment

Use the Hydra entrypoint in src.main:

uv run -m src.main problem=<problem> algorithm=<algorithm>

Example:

uv run -m src.main problem=sphere algorithm=mo_cma_mae

Supported problems and algorithms

Problems	Algorithms
sphere	mo_cma_mae
rastrigin	mome
arm	como_cma_es

Output

Experiments write logs under:

outputs/<problem>/<algorithm>/<timestamp>

The following artifacts are produced:

• summary.csv — QD metrics
• heatmap_*.png — archive heatmaps at logging intervals
• scheduler_*.pkl — checkpointed scheduler state

Configuration

The default experiment settings are defined in config/main.yaml. Algorithm and problem-specific settings are defined in config/algorithm/*.yaml and config/problem/*.yaml.

If resume_from_logdir is set in the Hydra config, the experiment will resume from the latest scheduler checkpoint in that directory.

TODO

• Add NSGA-II as a baseline algorithm
• Add SMS-EMOA as a baseline algorithm
• Add Overcooked as a benchmark problem