Enhanced methods for Evolution in-Materio Processors

Abstract

Evolution-in-Materio (EiM) is an unconventional computing paradigm, which uses an Evolutionary Algorithm (EA) to configure a material's parameters so that it can perform a computational task. While EiM processors show promise, slow manufacturing and physical experimentation hinder their development. Simulations based on a physical model were used to efficiently investigate three specific enhancements to EiM processors which operate as classifiers. Firstly, an adapted Differential Evolution algorithm that includes batching and a validation dataset. This allows more generational updates and a validation metric which could tune hyper-parameters. Secondly, the introduction of Binary Cross Entropy as an objective function for the EA, a continuous fitness metric with several advantages over the commonly used classification error objective function. Finally, the use of regression to quickly assess the material processor's output states and produce an optimal readout layer, a significant improvement over fixed or evolved interpretation schemes which can ‘hide’ the true performance of a material processor. Together these enhancements provide guidance on the production of more flexible, better performing, and robust EiM processors.