Lifelong control through Neuro-Evolution

ICLR 2026 Conference Submission20210 Authors

Keywords: evolutiion, neuro-evolution, reinforcement learning, lifelong learning

Abstract:

Reinforcement learning (RL) under continual environmental changes has remained a central challenge for decades.
Novel designs of loss functions, training procedures and neural network architectures have not yet managed to alleviate the main mode of failure in lifelong learning: loss of plasticity.
Here, we turn to a very different family of optimisers: neuro-evolution (NE).
Through an extensive evaluation on diverse lifelong control tasks, we see that both population-based and distribution-based approaches exhibit a remarkable ability to adapt where RL fails catastrophically.
We observe that, in the present of environmental shifts, NE naturally increases its diversity of solutions, evolving the ability to rapidly discover well-performing specialist individuals.
We propose that NE can be a promising approach towards tackling the need for lifelong adaptation and that future work should focus on the benefit of diversity.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 20210