STUDYING PROBABILISTIC FAULTS IN EVOLVED NONUNIFORM CELLULAR-AUTOMATA

We study the effects of random faults on the behavior of one-dimension al, non-uniform cellular automata (CA), where the local update rule ne ed not be identical for all grid sites. The CA systems examined were o btained via an approach known as cellular programming, which involves the evolution of non-uniform CAs to perform non-trivial computational tasks. Using the ''system replicas'' methodology, involving a comparis on between a perfect, non-perturbed version of the CA and a faulty one , we find that our evolved systems exhibit graceful degradation in per formance, able to tolerate a certain level of faults. We then ''zoom'' into the fault-tolerant zone, where ''good'' computational behavior i s exhibited, introducing measures to fine-tune our understanding of th e faulty CAs' operation. We study the error level as a function of tim e and space, as well as the recuperation time needed to recover from f aults. Our investigation reveals an intricate interplay between tempor al and spatial factors, with the presence of different rules in the gr id giving rise to complex dynamics. Studies along this line may have a pplications to future computing systems that will contain thousands or even millions of computing elements, rendering crucial the issue of r esilience.