Evolutionary learning in reaction-diffusion neurons

摘要

Reaction-diffusion neurons are neuron models in which synaptic inputs are processed by diffusion of second-messenger substances across a grid of fixed readout enzymes. This model is based upon the role of cyclic AMP in real neurons as an intraneuronal messenger, triggered by postsynaptic activation of cyclase enzymes, and exerting its effects by causing the phosphorylation of membrane proteins which regulate ion permeabilities. The firing behavior of such a neuron can be controlled by varying the distribution of kinase enzymes on the membrane. This suggests that a plausible neural learning mechanism involves the selection of nets of neurons with appropriately distributed kinases. We have developed a model of (ontogenetic) evolutionary learning which uses variation, propagation, and selection to generate optimal behaviors in small circuits of these cells. To test the capabilities of such a system, we simulate a simple motion learning task, in which the robot motor neurons are governed by internal reaction-diffusion dynamics, and we characterize the learnable behaviors. We argue that the neuronal dynamics of the cyclic nucleotide system contributes to the effectiveness of this evolutionary learning process.