2 ms·

Evolutionary Computation – A Practical Guide

• sevensor 9y ago
  The problem with the internet is that when you post about a subject you're not an expert in, some jerk with a PhD in the area is going to come in and rain on your parade. Today that jerk is me. I'm not going to comment too much on the deficiencies of modeling (physiology is hard), except to say that a game engine is probably going to spend time doing game-enginey things that aren't relevant to running your model. More evaluations is always better. The pure-mutation approach is fine, this is how Evolution Strategies work, but you might want to consider both (a) using a really good ES like CMAES and (b) double-checking that your solutions are actually better than random sampling. Alternatively (w.r.t CMAES, which is single-objective) you may want to consider using a multi-objective evolutionary algorithm like NSGA2. The article comments on the difficulty of choosing a fitness function. This is exacerbated by the need to choose only one fitness function. Multi-objective optimization is a natural fit for population-based search, because different population members can represent different points in objective space. You could for example have an objective for distance traveled and an objective for speed, to prevent slow, stable solutions from completely dominating. Finally, proportional selection is generally bad. Use tournament selection instead to avoid premature convergence.