This thesis illustrates the use and development of physical conductive analogue systems for unconventional computing using the Evolution in-Materio (EiM) paradigm. EiM uses an Evolutionary Algorithm to configure and exploit a physical material (or medium) for computation. While EiM processors show promise, fundamental questions and scaling issues remain. Additionally, their development is hindered by slow manufacturing and physical experimentation. This work addressed these issues by implementing simulated models to speed up research efforts, followed by investigations of physically implemented novel in-materio devices.

Initial work leveraged simulated conductive networks as single substrate ‘monolithic’ EiM processors, performing classification by formulating the system as an optimisation problem, solved using Differential Evolution. Different material properties and algorithm parameters were isolated and investigated; which explained the capabilities of configurable parameters and showed ideal nanomaterial choice depended upon problem complexity. Subsequently, drawing from concepts in the wider Machine Learning field, several enhancements to monolithic EiM processors were proposed and investigated. These ensured more efficient use of training data, better classification decision boundary placement, an independently optimised readout layer, and a smoother search space. Finally, scalability and performance issues were addressed by constructing in-Materio Neural Networks (iM-NNs), where several EiM processors were stacked in parallel and operated as physical realisations of Hidden Layer neurons. Greater flexibility in system implementation was achieved by re-using a single physical substrate recursively as several virtual neurons, but this sacrificed faster parallelised execution. These novel iM-NNs were first implemented using Simulated in-Materio neurons, and trained for classification as Extreme Learning Machines, which were found to outperform artificial networks of a similar size. Physical iM-NN were then implemented using a Raspberry Pi, custom Hardware Interface and Lambda Diode based Physical in-Materio neurons, which were trained successfully with neuroevolution. A more complex AutoEncoder structure was then proposed and implemented physically to perform dimensionality reduction on a handwritten digits dataset, outperforming both Principal Component Analysis and artificial AutoEncoders.

This work presents an approach to exploit systems with interesting physical dynamics, and leverage them as a computational resource. Such systems could become low power, high speed, unconventional computing assets in the future.