Abstract
This study investigates how dynamical systems may be learned and modelledwith a neuromorphic network which is itself a dynamical system. Theneuromorphic network used in this study is based on a complex electricalcircuit comprised of memristive elements that produce neuro-synaptic nonlinearresponses to input electrical signals. To determine how computation may beperformed using the physics of the underlying system, the neuromorphic networkwas simulated and evaluated on autonomous prediction of a multivariate chaotictime series, implemented with a reservoir computing framework. Throughmanipulating only input electrodes and voltages, optimal nonlinear dynamicalresponses were found when input voltages maximise the number of memristivecomponents whose internal dynamics explore the entire dynamical range of thememristor model. Increasing the network coverage with the input electrodes wasfound to suppress other nonlinear responses that are less conducive tolearning. These results provide valuable insights into how a practicalneuromorphic network device can be optimised for learning complex dynamicalsystems using only external control parameters.