NEOL framework proves hybrid neuroevolution achieves sublinear regret
First theoretical guarantee for combining architecture search with online weight adaptation
A new paper from researchers Shishen Lin and Yixin Chen, accepted at IJCAI-ECAI 2026, presents NEOL (NeuroEvolutionary Online Learning), a framework that splits neuroevolution into two timescales: an outer loop for architecture search (e.g., NEAT) and an inner loop for online weight adaptation using reward-modulated plasticity. This hybrid approach addresses the poor scaling of standard offline NEAT in high-dimensional control tasks. The key theoretical contribution is the first regret analysis for such a framework, proving that NEOL achieves sublinear regret—meaning performance improves predictably over time without hitting a plateau.
Empirically, the authors tested a NEAT-based NEOL implementation on four standard continuous control benchmarks. Under fixed interaction budgets, NEOL achieved significantly higher final fitness and lower variance compared to pure NEAT, and was competitive with state-of-the-art reinforcement learning baselines on several tasks. Statistical significance was confirmed via Wilcoxon rank-sum tests and ablation studies. The results suggest that incorporating online plasticity can dramatically improve the sample efficiency and robustness of neuroevolution, making it a practical alternative to RL in resource-constrained settings. Code is available online.
- NEOL provably achieves sublinear regret, the first theoretical guarantee for hybrid neuroevolution with online learning
- On four control benchmarks, NEOL outperforms pure NEAT in final fitness and variance under fixed interaction budgets
- Competitive with strong RL baselines, with code released and accepted at IJCAI-ECAI 2026
Why It Matters
Bridges neuroevolution and online learning for more efficient, provably convergent architecture search in continuous control.