Reservoir computing with networks of nanoscale memristors: optimisation of memristor responses for maximal computational performance

Abstract

Self-assembled networks of memristive elements have been shown to be effective in several types of computation, including reservoir computing. Here we use realistic simulations to explore the maximal performance that is possible in several standard computational tasks. Specifically, we investigate the extent to which tuning the parameters that control the properties of the individual memristors improves computational performance of the networks, and show that for some computational tasks a significant improvement (a factor of two) is possible with respect to previous results. We also discuss the effects of noise and compare with hypothetical devices in which outputs can be measured from all nodes of the network, and show that performance of realistic devices with a finite number of electrodes is not dramatically lower.