Evolutionary optimization reveals structural rules for reservoir computing chaos prediction

Reservoir computing traditionally treats the recurrent network as a fixed random substrate and only trains the readout. Nima Dehghani flipped this by placing the entire reservoir architecture under evolutionary selection for predicting spatiotemporal chaos (the Kuramoto–Sivashinsky equation). Five construction hyperparameters—size, connectivity degree, spectral radius, input scaling, and readout regularization—were optimized across generations.

Evolution drove down prediction error at the population level and extended the low-error forecast horizon. Structural analysis revealed that evolved reservoirs stabilized within a conserved stochastic-block-model-like spectral envelope while refining low-eigenvalue modes. Modularity was locked to an intermediate band, and connection cost was pruned within that band. Pareto analysis showed elite reservoirs formed a horizontal floor in the cost–modularity plane, meaning accuracy and efficiency were achieved jointly rather than through a simple trade-off.

These findings demonstrate that evolutionary optimization doesn't just improve performance—it exposes interpretable structural constraints on the recurrent substrate. The work provides a bio-inspired framework for understanding how predictive demands shape adaptive dynamical networks, with implications for designing efficient, task-tuned reservoir computers for complex physical systems.