Reservoir Computing in Materio:An Evaluation of Configuration through Evolution

Recent work has shown that computational substrates made from carbon nanotube/polymer mixtures can form trainable Reservoir Computers. This new reservoir computing platform uses computer based evolutionary algorithms to optimise a set of electrical control signals to induce reservoir properties within the substrate. In the training process, evolution decides the value of analogue control signals (voltages) and the location of inputs and outputs on the substrate that improve the performance of the subsequently trained reservoir readout. Here, we evaluate the performance of evolutionary search compared to randomly assigned electrical configurations. The substrate is trained and evaluated on time-series prediction using the Santa Fe Laser generated competition data (dataset A). In addition to the main investigation, we introduce two new features closely linked to the traditional reservoir computing architecture, adding an evolvable input-weighting mechanism and a reservoir time-scaling parameter. The experimental results show evolved configurations across all four test substrates consistently produce reservoirs with greater performance than randomly configured reservoirs. The results also show that applying both input-weighting and timescaling simultaneously can provide additional tuning to the task, improving performance. For one material, the evolved reservoir is shown to outperform – for this task – all other hardwarebased reservoir computers found in the literature. The same material also outperforms a simple evolved simulated Echo State Network of the same size. The performance of this material is reported to be both consistent after long time-periods and after reconfiguration to other tasks.