Research map / Reinforcement learning / Visual and representation RL
ConfigX: Modular Configuration for Evolutionary Algorithms via Multitask Reinforcement Learning
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 ConfigX, a Transformer-based universal configuration policy for evolutionary algorithms, meta-trained with multitask PPO over modularly generated EA structures (Modular-BBO) and problem instances, showing zero-shot generalization and fine-tuning adaptation.
Problem
Automatic configuration of evolutionary algorithms is usually tailored to a specific EA and problem, requiring retraining or redesign for each; the paper asks whether one meta-learned agent can configure diverse EAs across problems.
Approach
A modular EA framework (Modular-BBO) with hierarchical polymorphism spans millions of algorithm structures; a Transformer policy conditioned on per-sub-module ID and performance tokens is trained with multitask PPO over joint algorithm-problem tasks, so one policy configures many EAs.
Key results
- On in-distribution test tasks, pre-trained ConfigX (9.81E-01) surpasses the SMAC3 baseline in zero-shot performance and clearly beats the Random baseline.
- On out-of-distribution problems (real-world Protein-docking/HPO-B) ConfigX achieves 9.86E-01 zero-shot performance.
- Fine-tuning saves 3x learning steps versus re-training on new problems and 2x on new algorithm structures.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace MSA blocks with an MLP layer (ablation) | ConfigX · T_test,in, T_test,out^(1), and T_test,out^(2) | normalized accumulated optimization improvement: 9.81E-01 -> 9.70E-01; 9.86E-01 -> 9.80E-01; 9.22E-01 -> 9.16E-01 | Hurt |
| remove sinusoidal positional embeddings (ablation) | ConfigX · T_test,in, T_test,out^(1), and T_test,out^(2) | normalized accumulated optimization improvement: 9.81E-01 -> 9.74E-01; 9.86E-01 -> 9.81E-01; 9.22E-01 -> 9.19E-01 | Hurt |
| replace sinusoidal positional embeddings with learnable positional embeddings (ablation) | ConfigX · T_test,in, T_test,out^(1), and T_test,out^(2) | normalized accumulated optimization improvement: 9.81E-01 -> 9.82E-01; 9.86E-01 -> 9.84E-01; 9.22E-01 -> 9.20E-01 | No clear effect |
| replace MSA blocks with an MLP layer and remove sinusoidal positional embeddings (ablation) | ConfigX · T_test,in, T_test,out^(1), and T_test,out^(2) | normalized accumulated optimization improvement: 9.81E-01 -> 9.51E-01; 9.86E-01 -> 9.73E-01; 9.22E-01 -> 9.06E-01 | Hurt |
Surprising findings
- Zero-shot performance on unseen algorithm structures (GA/PSO trained only on DE) is not as expected, showing a generalization gap across EA families.
Limitations
- Zero-shot performance on unseen algorithm families (GA/PSO) is weaker, requiring fine-tuning.
- Pre-trained model cannot cover sub-modules absent from its training algorithm space.
- Training task set is limited to DE algorithm structures for the main training run.
Experimental setup
Datasets: CoCo-BBOB Protein-docking benchmark HPO-B
Benchmarks: CoCo-BBOB Protein-docking benchmark HPO-B
Metrics: normalized accumulated optimization improvement
Baselines: SMAC3 Original Random ConfigX-MLP ConfigX-NPE ConfigX-LPE MLP-NPE
Builds on: PPO MetaBBO SMAC3
Closest papers by meaning
- MetaBox-v2: A Unified Benchmark Platform for Meta-Black-Box Optimization (NeurIPS 2025 · similarity 0.90)
- DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization (NeurIPS 2025 · similarity 0.88)
- Evolution-Guided Policy Gradient in Reinforcement Learning (NeurIPS 2018 · similarity 0.88)
- $\texttt{MetaDistill}$: Unlocking the Performance Ceiling for Pretrained Optimizers (ICML 2026 · similarity 0.87)
- MetaBox: A Benchmark Platform for Meta-Black-Box Optimization with Reinforcement Learning (NeurIPS 2023 · similarity 0.87)
- Graph-Supported Dynamic Algorithm Configuration for Multi-Objective Combinatorial Optimization (ICML 2025 · similarity 0.86)
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.