Is Evolution Predictable? Experiments in an Evolutionary Video Game

Abstract

The outcome of evolution sometimes appears to be predictable, as in the evolution of the same characteristics independently in convergent evolution. Other times, evolutions path depends on starting conditions or chance events, and some forms evolve just once and never appear again. Is convergence, and by implication, predictability, a common characteristic of the evolutionary process? We aimed to answer this question by using an evolutionary video game titled Project Hastur as a study system. In this video game, the enemies evolve traits to help them combat the players strategy. We determined whether the same environmental pressures (in this case, player strategy) lead to predictable evolution. We conducted a series of experiments with three different playing styles and four different evolutionary treatments with varying kinds of selection pressure, each with several replicates. For each replicate, enemies from the first and final generations were categorized into types using k-means clustering. The Euclidean distance between the cluster centroids and the sum of squares values were recorded for each replicate and compared among evolutionary treatments using Wilcoxon rank sum tests. Evolutionary predictability was evaluated using permutation tests. We found that fitness functions in Project Hastur led to incredible, but unpredictable, diversity. The fitness landscape changes between replicates, even within the same experimental treatment and regardless of player strategy, resulting in enemies with an unpredictable array of trait values.