EvoForest: Open-Ended Evolution of Computational Graphs for ML
A new machine-learning paradigm called EvoForest has been introduced in a paper on arXiv (2604.19761). It proposes open-ended evolution of computational graphs as an alternative to traditional parameter optimization. The system jointly evolves reusable computational structures, callable function families, and trainable continuous components, targeting structured prediction problems where discovering what to compute is the bottleneck. This hybrid neuro-symbolic approach addresses non-differentiable objectives, cross-validation-based evaluation, interpretability, and continual adaptation.
Key facts
- EvoForest is a hybrid neuro-symbolic system for end-to-end open-ended evolution of computation.
- It jointly evolves reusable computational structure, callable function families, and trainable low-dimensional continuous components.
- The paper is published on arXiv with ID 2604.19761.
- It targets structured prediction problems where the main bottleneck is discovering what should be computed from the data.
- Traditional ML focuses on choosing a parameterized model family and optimizing its weights.
- EvoForest addresses non-differentiable objectives, cross-validation-based evaluation, interpretability, and continual adaptation.
- The system evolves computational graphs rather than merely generating features.
- The approach is described as a novel machine-learning paradigm.
Entities
Institutions
- arXiv