The Application of a Virtual Programmable Logic Device for Robotic Control and Pattern Recognition

Abstract

This research derives and evaluates a novel machine learning architecture called the Virtual Programmable Logic Device (VPLD), and if the VPLD can become a competitor to the Artificial Neural Network (ANN) when evolved for applications in robotic control and pattern recognition. The VPLD is based on the architecture of a programmable logic device but is coded in software rather than in hardware. This allows the VPLD to be run on CPU based platforms including standard PCs, mobile phones and ARM based embedded systems such as the Raspberry Pi. The operation of the VPLD can be configured and optimised using evolutionary algorithms. The VPLD is inspired by previous Evolvable Hardware architectures evolved for applications such as robotic control. In the Evolvable Hardware domain electronic circuits are evolved on programmable logic devices such as the field programmable gate array (FPGA).

The VPLD is investigated in two fields: 1) evolutionary robotics where the gait control of a hexapod robot and the autonomous navigation of a two-wheel drive mobile robot is examined; and 2) in pattern recognition where character recognition, and melanoma classification are evaluated. In both application domains two types of VPLD are investigated, the first is the digital VPLD (D-VPLD) which mimics the PLD binary variables and digital logic, the second is the floating-point VPLD (F-VPLD) which uses floating-point variables and mathematical functions. The floating-point variables are complex to implement in hardware on a FPGA.

In the gait control of a hexapod, a evolvable hardware implementation is designed to validate the VPLD architecture. In this validation it is demonstrated that the VPLD compared to the evolvable hardware for robotic control is faster to evolve, as well as simpler and cheaper to implement.

To assess the VPLDs controller and classifier performance it is benchmarked against an ANN. The VPLD and ANN are evolved for the robotic control and pattern recognition problems using the same evolutionary algorithm. The results of the experiments show the VPLD is a viable alternative to the ANN in both robotic control and pattern recognition applications as the VPLD could achieve the same, and in some cases better performance than the ANN.