Results

No parasites or similar phenomena were observed in any of the nine runs. It appears that the mechanism used to simulate the reading of neighbouring programs' code was too complicated to be exploited by evolution. In particular, unlike in Tierra (see Section 3.3.1), there is no single mutation of the ancestor which would produce a functional parasite.

An analysis of the activity of the most numerous genotypes in each of the runs shows that they generally produced a much larger proportion of unfaithful offspring than did genotypes in the standard run. For example, the data recorded in the output file morgue.dat shows that the mean number of unfaithful reproductions for programs of the ancestor genotype 348AAAA was 0.277 (with standard deviation 0.099) across all nine runs. This compares to a mean of 0.010 unfaithful reproductions (with standard deviation 0.003) in the nine re-runs of the standard model, in which neighbouring code was not readable.^6.8

This decrease in the copy fidelity of the programs was evidently caused by the mechanism for reading neighbouring code. Recall that this mechanism will attempt to bind a promoter onto the code of every available neighbour before it tries the host code. In eight of the nine runs, a state resembling Class 1 dynamics quickly arose, with the appearance of programs with length of order 1000 bits, and population sizes of order 400 individuals. The high proportional of unfaithful offspring produced by programs meant that the evolutionary potential of the runs was limited right from the start, and no trends for adaptive improvements of any sort were observed. In the other run (Run 7), the population was reduced to a single program by time slice 87,600, which survived until the end of the run.

In summary, the mechanism for reading the code of neighbouring programs failed to reproduce the results observed in Tierra. In particular, no parasites emerged during these runs.