Hyper Algorithm Design Agent (HADA)

This repo provides the implementation of HADA (Hyper Algorithm Design Agent). HADA automates the design of Meta-Black-Box Optimization (MetaBBO) algorithms through open-ended recursive program evolution.

HADA's paper is available here. Consider cite it if necessary.

@article{yu2026hyper,
  title={Hyper Algorithm Design Agent: Evolving Learnable Optimizer from Zero},
  author={Yu, Zipei and Gong, Yue-Jiao and Ma, Zeyuan and Jiang, Yuncheng and Cao, Zhiguang},
  journal={arXiv preprint arXiv:2609.35328},
  year={2026}
}

We provide following step-by-step instructions for you to try HADA on your PC or Server. Raise issues if any problem happens during the using.

Quick Start

0. Prepare a ubuntu system (empirically, the version of the ubuntu doesn't matter), we use 24.04.4 LTS in our paper.

1. Prepare your API Key

Fill in your API Key in the env.txt file and rename it to .env

2. Create virtual environment

python3 -m venv venv_nat
source venv_nat/bin/activate
pip install -r requirements.txt

3. Build Docker Image

docker build -t hada -f Dockerfile .

4. Initialize Git Repository

git init
git add .
git commit -m "Initial commit"

5. Run Initial Evaluation

Run initial evaluation to establish baseline performance:

# Run all domains
bash setup_initial.sh

# Or run specific domain only
bash setup_initial.sh bbob_unconstrained
bash setup_initial.sh bbob_constrained
bash setup_initial.sh metabox_mo

An initial evluation output folder will show up in ./outputs/ after the above operation.

6. Start HADA Generation Loop

python generate_loop.py --domains bbob_unconstrained --max_generation 100

Supported Domains

• bbob_unconstrained: BBOB unconstrained optimization problems
• bbob_constrained: BBOB constrained optimization problems
• metabox_mo: WFG multi-objective optimization problems

Project Structure

HADA/
├── Dockerfile              # Docker image build file
├── setup_initial.sh        # Initial evaluation script
├── generate_loop.py        # Main generation loop entry
├── metabbo/                # DQN-DE core implementation
│   ├── ec_algorithm.py     # DE optimizer
│   ├── meta_learning.py    # DQN controller
│   └── meta_learning_env.py # RL environment wrapper
├── domains/                # Evaluation logic for each domain
│   ├── bbob_unconstrained/
│   ├── bbob_constrained/
│   └── metabox_mo/
└── outputs/                # Evaluation output directory

Main Parameters

generate_loop.py supports the following parameters:

• --domains: Domains to optimize (comma-separated)
• --max_generation: Maximum number of generations
• --parent_selection: Parent selection strategy (score_prop/score_child_prop/random/latest/best)

Output

For each evolution step, HADA generates a in-execution information folder (named by gen_xx) in the outputs/ directory, which includes:

• hyper_agent_chat_history.md: the LLM conversation history for hyper agent.
• chat_history_task_agent.md: the LLM conversation history for task agent.
• hyper_agent_patch.diff: The patches made at current step in hyper agent.
• all_patch.diff: The patches made at current step in task agent and metabbo.
• report.json: Evaluation report
• predictions.csv: The optimization results (performance score) of the current metabbo on the target domain.
• {domain}_return_curves.png: Training return curves of the metabbo.
• {domain}_loss_curves.png: Training loss curves of the metabbo.

There are also two important experimental progress file you need to pay attention, they are dynamically updated along HADA's running:

• archive_graph_bbob_unconstrained_train_agent.png: the algorithm evolution tree updated at each step.
• progress_plot_bbob_unconstrained_train_agent.png: the performance progress curve of HADA-evolved metabbo updated at each step.