Research map / Reinforcement learning / Visual and representation RL
MetaBox-v2: A Unified Benchmark Platform for Meta-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.
Presents MetaBox-v2, a unified open-source platform for meta-black-box optimization that supports RL, SL, NE and ICL paradigms, adds parallel training/testing, 18 testsuites, and new metrics, and benchmarks 20 baselines.
Problem
Existing MetaBBO benchmark (MetaBox 2023) only supports single-objective, RL-based MetaBBO with serial evaluation, which no longer covers the rapidly expanding MetaBBO paradigms and optimization scopes.
Approach
A unified Basic_Agent interface wraps RL, gradient-based (SL) and evolutionary (NE) learning objectives plus LLM in-context designs under one MetaBBO template, with a polymorphic Basic_Problem class for diverse optimization scopes, vectorized training environments, Ray-based parallel testing, and two new metrics (learning efficiency and Anti-NFL).
Key results
- Parallel training and testing accelerate MetaBBO by 10-40x; testing Mode-4 gives no less than 40x throughput over original MetaBox.
- DEDDQN ranks first on average in the bbob-10D in-distribution test, winning 6 of 18 instances, while MetaBBO baselines win most instances overall.
- No single baseline dominates all others in the learning-efficiency vs. Anti-NFL trade-off; RLDAS is efficient but has among the lowest Anti-NFL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| parallelize training with a vectorized optimization environment | MetaBox-v2 · MetaBBO training | speedup: at most 10x | Helped |
| distribute all evaluation tasks with Mode-4 | MetaBox-v2 testing · MetaBBO testing | speedup: no less than 40x | Helped |
| use MetaBBO baselines instead of traditional BBO baselines | bbob-10D testing · bbob-10D testing instances | optimization performance | Helped |
| deploy DEDDQN on the protein suite after training on bbob-10D | DEDDQN · protein | optimization performance | Hurt |
| use a Transformer meta-policy rather than MLP or LSTM policies | MetaBBO on high-dimensional neuroevolution tasks · ne (Ant-3, Ant-4, and Ant-5) | optimization performance | Helped |
| use RLDAS rather than less efficient MetaBBO baselines | MetaBBO baselines · all testsuites | learning efficiency | Helped |
| train DEDDQN for its best performance | DEDDQN · all testsuites | training efficiency: hundreds of hours | Hurt |
| use GLHF or LES for cross-suite robustness | MetaBBO baselines · unseen testsuites | Anti-NFL | Helped |
| use RLDAS across different problem dimensions | RLDAS · unseen testsuites | Anti-NFL | Hurt |
Surprising findings
- Traditional BBO optimizers such as DE and SHADE show more robust out-of-distribution performance than many MetaBBO baselines, which appear to overestimate generalization.
- DEDDQN performs well in-distribution but reverses on the protein testsuite, indicating overfitting.
Limitations
- Case study covers only 20 selected baselines and a limited set of out-of-distribution testsuites.
- Anti-NFL and learning-efficiency metrics are proposed without consensus on their standardization; high Anti-NFL baselines show low absolute performance.
Experimental setup
Datasets: bbob-10D bbob-noisy-30D protein uav hpo-b bbob-30D lsgo ne zdt uf dtlz wfg
Benchmarks: bbob-10D bbob-noisy-30D protein uav hpo-b ne
Metrics: normalized performance throughput (instance runs per second) learning efficiency Anti-NFL indicator average rank
Baselines: PSO DE SHADE JDE21 MadDE RNNOPT DEDDQN DEDQN LDE RLPSO RLEPSO NRLPSO
Builds on: MetaBox Tianshou Ray DEAP
Closest papers by meaning
- ConfigX: Modular Configuration for Evolutionary Algorithms via Multitask Reinforcement Learning (AAAI 2025 · similarity 0.90)
- DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization (NeurIPS 2025 · similarity 0.90)
- Meta-Black-Box-Optimization through Offline Q-function Learning (ICML 2025 · similarity 0.89)
- SYMBOL: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning (ICLR 2024 · similarity 0.89)
- SMOG: Scalable Meta-Learning for Multi-Objective Bayesian Optimization (NeurIPS 2026 · similarity 0.88)
- Projection Optimization: A General Framework for Multi-Objective and Multi-Group RLHF (ICML 2025 · similarity 0.88)
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.