Building multimodule systems with unlimited evolvable capacities from modules with limited evolvable capacities (MECs)

Abstract

This paper introduces a concept which we believe will play a fundamental role in the growing field of "evolutionary engineering", namely the idea that there are limits to what can be evolved using a finite number of bits in a chromosome. For example, if one tries to evolve a neural network circuit module to give a time varying analog output signal which tracks an analog output time varying target signal, then the actual evolved output curve will follow the target curve quite well for a certain time period, then diverge. If one puts more bits into the chromosome used to evolve the signal, then The evolved signal will track the target signal for longer, but again will eventually diverge. Hence there is a finite "evolvable capacity" for a module evolved with a given number of bits. We label this concept "modular evolvable capacity" or simply MEC. MECs are important when one attempts to assemble large numbers of evolved modules to build such systems as artificial brains. STARLAB will attempt to use its CAM-Brain Machine (CBM) to evolve and assemble 64000 such modules to build an artificial brain. The fact that each module has its MEC, places constraints upon what "evolutionary engineers (EEs)", or in this case "brain architects (BAs)" can do. Such limits are unavoidable and have a fundamental practical impact on the daily work of EEs and BAs. This paper aims to show how multimodule systems with effectively unlimited evolvable capacities may be buildable using modules with limited MECs.