Creating Unorganised Machines from Memristors

摘要

There is growing interest in memristive devices following their recent nanoscale fabrication. This paper describes initial consideration of the implementation of artificial intelligence within predominantly memristive hardware. In particular, versions of Alan Turing’s discrete dynamical network formalism — the unorganised machine — are used as the knowledge representation scheme and a population-based search technique is used to design appropriate networks. Issues including memristor count and global network synchrony are compared for two memristive logic implementations (NAND and IMP) on a well-known simulated robotics benchmark task. It is shown that IMP networks are harder to design than NAND, but are simpler to implement and require fewer processor cycles.