Virtual Creatures Evolved Using Speciation and Historical Markings

The design of virtual creatures shown in video [3] improves on the original design of Sims [4] by using the concept of historical markings for tracking the ancestry of every body part and every neuron of the creature throughout the evolution (a concept inspired by NEAT [5], an algorithm for evolution of neural networks). Historical markings are unique identifiers assigned to every body part and every neuron of the creature’s genotype at the moment when the part is created by mutation. Markings are also inherited during recombination. This makes it possible to take any two creatures in the population and (1) quantify in a reliable way how similar they are to each other (based on what proportion of their bodies and neural networks shares the same historical markings, i.e. come from the same common ancestor) and (2) design a robust recombination algorithm by swapping properties of body parts or neurons that have matching values of historical markings. Being able to reliably define similarity of any two creatures then makes it possible to effectively implement speciation as a way of maintaining diversity in the population. Speciation is implemented by dividing all creatures from the current population into groups (species) based on their similarity to each other. Creatures compete and mate only with creatures in the same species. The number of offspring of each species is proportional to the average fitness of all creatures in that species. This allows new body parts and new neurons (which might initially be disadvantageous to a creature) to improve in a separate species, instead of being immediately dominated by currently better creatures in the entire population. Large-scale experiments presented in [1, 2] have shown that the new algorithm requires smaller number of fitness evaluations to evolve creatures with a given fitness value than the standard algorithm used by Sims [4] (where no speciation is used and grafting and crossover are used instead of markings-based recombination), and have also confirmed that every component of the algorithm is also beneficial when used on its own.