Evolutionary Emergence of Neurodynamic Networks for Robust Control: A Simple Excitatory-Inhibitory Network

ICLR 2026 Conference Submission15374 Authors

Keywords: Bio-inspired learning, evolutionary computation, reinforcement learning, neurodynamic networks

TL;DR: Using EA to train an EI network from a random state to an organized functional structure to perform reinforcement learning tasks

Abstract: Fine-grained network models based on differential equations, and neurodynamic synapses and neurons provide a realistic description of biological neuronal networks, compared with mainstream artificial neural networks. They nevertheless have not been widely explored, mainly due to the lack of effective parameter training methods. We propose a neurodynamic model training method that combines an efficient neurodynamic simulation architecture and an evolutionary algorithm. Based on a simple Excitatory-Inhibitory network, a neurodynamic model with task control capabilities is successfully obtained via parallel dynamic simulation, and network selection methods under evolutionary pressure. Compared with the state-of-the-art reinforcement learning methods, the resulting neurodynamic network can achieve comparable task control performance for Mojoco tasks in a significantly smaller network scale within fewer training steps. Our work provides an alternative path to functional networks alongside mainstream reinforcement learning frameworks, and prove the feasibility of the evolutionary approach toward biological intelligence.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 15374