Machine Learning

Algorithmically probable mutations reproduce aspects of evolution such as convergence rate, genetic memory, modularity, diversity explosions, and mass extinction

This topic contains 0 replies, has 1 voice, and was last updated by  arXiv 1 year, 3 months ago.


  • arXiv
    5 pts

    Algorithmically probable mutations reproduce aspects of evolution such as convergence rate, genetic memory, modularity, diversity explosions, and mass extinction

    We show that if evolution is algorithmic in any form and can thus be considered a program in software space, the emergence of a natural algorithmic probability distribution has the potential to become an accelerating mechanism. When the strategy produces unfit organisms massive extinctions occur and modularity provides an evolutionary advantage evolving a genetic memory. We simulate the application of these mutations (no recombination) based on numerical approximations to algorithmic probability, in what constitutes the first experiments in artificial life based on aspects of Algorithmic Information Theory. We find that recurring structures can rapidly become pervasive, potentially explaining some aspects of convergent evolution, and that the emergence of information modules by local evolution is unavoidable, requiring memory support reminiscent of functional structures such as genes and biological information carriers such as DNA. We demonstrate that such regular structures are preserved and carried on when they first occur and can also lead to an accelerated production of diversity and extinction, possibly explaining natural phenomena such as periods of accelerated growth of the number of species and the occurrence of massive extinctions whose causes are a matter of considerable debate. The approach introduced here appears to be a better approximation to actual biological evolution than models based upon the application of mutation from uniform probability distributions, and because evolution by algorithmic probability converges faster to regular structures (both artificial and natural, as tested on a small biological network), it also approaches a formal version of open-ended evolution based on previous results. We also show that the procedure has the potential to significantly accelerate solving optimization problems in the context of artificial evolutionary algorithms.

    Algorithmically probable mutations reproduce aspects of evolution such as convergence rate, genetic memory, modularity, diversity explosions, and mass extinction
    by Santiago Hernández-Orozco, Hector Zenil, Narsis A Kiani
    https://arxiv.org/pdf/1709.00268v3.pdf

You must be logged in to reply to this topic.