The Virtual Programmable Logic Device, a Novel Machine Learning Architecture

This paper introduces a novel architecture for robotic control, called a Virtual Programmable Logic Device (VPLD) where the operation of the device can be dynamically configured using machine learning methods. The VPLD is based on the structure of a Programable Logic Device (PLD) comprising of a two-dimensional feed-forward array of function blocks, with each block containing multiplexers and function elements. However, the VPLD is implemented in software rather than hardware allowing the VPLD to be run on a CPU based system, such as PC's and ARM based embedded systems. The operation of the VPLD is determined by a configuration bitstream which configures the multiplexers (routing) and function elements of each block. A genetic algorithm is used to evolve the configuration bitstream to produce the walking gait of a hexapod robot. The controller performance and the evolutionary efficiency of the VPLD are compared with an evolved artificial neural network (ANN), and an evolvable hardware (EHW) device. It was found the VPLD and ANN had similar controller performance and evolutionary efficiency, while the EHW had a comparable controller performance however its evolutionary efficiency was poor. It is shown that the VPLD is a viable alternative to an ANN for evolutionary robotic control.