Research map / Probabilistic and scientific modeling / Bayesian optimization
DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DesignX, a dual-agent Transformer RL framework that jointly generates optimizer workflows (Agent-1, REINFORCE) and controls hyperparameters (Agent-2, PPO) over an extended modular EC space, meta-trained on 10k synthetic BBO problems.
Problem
Automating black-box optimizer design requires both choosing an algorithm workflow and tuning its hyperparameters, but prior MetaBBO methods address only one of these sub-tasks, yielding suboptimal optimizers.
Approach
A dual-agent system where a GPT-2-based workflow generator (Agent-1) and a GPT-2-based hyperparameter controller (Agent-2) are trained jointly with a cooperative reward over a modular EC algorithm space (Modular-EC, 116 modules across DE/ES/GA/PSO), with legality ensured by topology-rule masking.
Key results
- On 20 sampled in-distribution test instances, DesignX achieves the best mean objective on most instances and the best normalized averaged objective (0.0826 vs 0.132 for the best MetaBBO baseline).
- DesignX generalizes out of distribution to Protein-Docking, HPO-B and UAV suites, outperforming human-crafted optimizers from several decades and MetaBBO baselines in average optimization curves.
- Ablation shows workflow generation (Agent-1) matters more than hyperparameter control (Agent-2), and joint training beats single-agent variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Jointly train workflow generation and dynamic hyperparameter control with the dual-agent system | DesignX · D_test (20-instance normalized average) | normalized averaged objective: 1.32E-01 ±7.36E-01 -> 8.26E-02 ±1.75E-01 | Helped |
| Use the DesignX-generated optimizer instead of the strongest listed baseline | F1 (MAH, 50D, 30000 FEs) · F1 | objective: 1.27E+00 ±4.41E-01 -> 2.89E-01 ±3.93E-01 | Helped |
| Use the DesignX-generated optimizer instead of MetaBBO | F154 (UAH, 50D, 10000 FEs) · F154 | objective: 5.47E+02 ±0.00E+00 -> 6.99E+02 ±7.45E+01 | Hurt |
| Train both agents rather than only Agent-1 or only Agent-2 (ablation) | DesignX · D_test and three realistic problem sets | reversed normalized objective | Helped |
| Train Agent-1 for workflow generation rather than using randomized Agent-1 (ablation) | DesignX without Agent-2 training · D_test and three realistic problem sets | reversed normalized objective | Helped |
| Generate customized workflows for each problem rather than use a static workflow (ablation) | DesignX (SBS ablation) · D_test and three realistic problem sets | reversed normalized objective | Helped |
| Use larger models as training-problem scale increases (ablation) | DesignX · D_test | average learning objective | Mixed |
| Replace problem and observation features with zero vectors (ablation) | DesignX (SBS) · D_test and three realistic problem sets | reversed normalized objective | Hurt |
Surprising findings
- Generating the correct workflow appears more important than hyperparameter control, per ablation comparisons.
- Larger model variants overfit when the training set is small, so a single-block GPT-2 was selected as final.
Limitations
- Restricted to single-objective optimization.
- Training takes about 6 days, which is inefficient.
- Relies on a hand-built modular algorithm space and 9 ELA features.
Experimental setup
Datasets: synthetic BBO problem set built from 32 basic functions (CEC-style) Protein-Docking HPO-B UAV path planning suite
Benchmarks: CEC-style synthetic test set (3200 instances) Protein-Docking HPO-B UAV path planning
Metrics: objective value (mean over 51 runs) normalized averaged objective critical difference rank
Baselines: GLHF DEDQN GLEET GA PSO DE CMA-ES SHADE
Builds on: ConfigX GLHF Modular-BBO REINFORCE PPO
Closest papers by meaning
- MetaBox-v2: A Unified Benchmark Platform for Meta-Black-Box Optimization (NeurIPS 2025 · similarity 0.90)
- ConfigX: Modular Configuration for Evolutionary Algorithms via Multitask Reinforcement Learning (AAAI 2025 · similarity 0.88)
- Teacher-Aware Evolution of Heuristic Programs from Learned Optimization Policies (NeurIPS 2026 · similarity 0.85)
- Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering with Generative Optimization (NeurIPS 2026 · similarity 0.85)
- SYMBOL: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning (ICLR 2024 · similarity 0.85)
- Decoupling Exploration and Policy Optimization: Uncertainty Guided Tree Search for Hard Exploration (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.