A Massively Parallel Architecture for Linear Machine Code Genetic Programming

Abstract

Over the last decades Genetic Algorithms (GA) and Genetic Programming (GP) have proven to be efficient tools for a wide range of applications. However, in order to solve human-competitive problems they require large amounts of computational power, particularly during fitness calculations.

In this paper I propose the implementation of a massively parallel model in hardware in order to speed up GP. This fine-grained diffusion architecture has the advantage over the popular Island model of being VLSI-friendly and is therefore small and portable, without sacrificing scalability and effectiveness. The diffusion architecture consists of a large amount of independent processing nodes, connected through an X-net topology, that evolve a large number of small, overlapping sub-populations. Every node has its own embedded CPU, which executes a linear machine code representation of the individuals. Preliminary simulation results (low-level VHDL simulation) indicate a performance of 10.000 generations per second (depending on the application). One node requires 10-20.000 gates including the CPU (also application dependent), which makes it possible to fit up to 2.000 individuals in one FPGA (Virtex XC2V10000).