Initial evolvability experiments on the CAM-brain machines (CBMs)

摘要

Presents the results of some of the first evolvability experiments undertaken on CAM (content-addressable memory) brain machines (CBMs), using the hardware itself (not software simulations). A CBM is a specialised piece of programmable (evolvable) hardware that uses Xilinx XC6264 programmable FPGA chips to grow and evolve, at electronic speeds, 3D cellular automata (CA) based neural network circuit modules of some 1,000 neurons each. A complete run of a genetic algorithm (e.g. with 100 generations and a population size of 100) is executed in a few seconds. 64,000 of these modules can be evolved separately according to the fitness definitions of human evolutionary engineers and downloaded, one by one, into a gigabyte of RAM. Human brain architects then interconnect these modules "by hand" according to their artificial brain architectures. The CBM then updates the binary neural signaling of the artificial brain (with 64,000 "hand" interconnected modules, i.e. 75 million neurons) at a rate of 130 billion CA cell updates a second, which is fast enough for the real-time control of robots. Before such multi-module artificial brains can be constructed, it is essential that the quality of the evolution (the "evolvability") of the individual modules should be adequate. This paper reports on the initial evolvability results obtained on CBM hardware.