For many years researchers inspired by the idea of natural selection have experimented with computer programs called evolutionary algorithms. These algorithms simulate a kind of artificial breeding process in which a set of candidates generated by the computer are evaluated for their ability to perform a desired task. The best performers are then allowed to reproduce with slight variation to form a new and hopefully improved generation. In recent years evolutionary algorithms have exhibited the ability to evolve brain-like structures called artificial neural networks in an approach called neuroevolution. These evolved networks perform tasks often critical to technological progress and artificial intelligence like controlling robots or recognizing images. However, unlike evolution in nature, which yields dramatic changes over hundreds of thousands of generations, evolutionary algorithms have rarely been run for more than a few thousand. This project for the first time is applying new evolutionary techniques that reward continual novelty and diversification to experiments evolving over hundreds of thousands of generations, on the scale of nature. The driving hypothesis is that modern evolutionary algorithms run on this scale can yield robotic behaviors, agent morphologies, and decision-making capabilities significantly beyond the current state of the art.
To investigate long-term evolution in practice, artificial neural networks are being evolved in a variety of domains through new kinds of novelty-driven neuroevolution algorithms designed to avoid the convergence seen in typical evolutionary experiments. Because this new class of algorithms tends to avoid convergence, the long-term dynamics and ultimate potential for discovery of such algorithms over vast time scales (i.e. hundreds of thousands of generations) is almost entirely unknown. The idea of running neuroevolution at unprecedented timescales mirrors recent results in related areas like deep learning where massive computation has proven capable of fundamentally altering the kinds of problems that can be solved. Because evolutionary runs over hundreds of thousands of generations yield enormous troves of evolutionary data, an important component of the project is the development visualization techniques for exploring and characterizing the results of such runs. The project overall is producing a new set of tools, a new set of evolved capabilities for autonomous control and decision-making, and an increased understanding of the implications of big data and big computation for simulated evolution in computers.
