its often the case that if you throw {evolution, gradient decent, simulated annealing, etc} at some optimization problem and your fitness function is ill specified, you get solutions which satisfy the fitness function but are “stupid” (ex: ask the algorithm to evolve creatures that can walk, you often get very tall creatures that fall over, because that solution is “easier” than actually getting the creature to walk). it seems like that this means you get “the first solution that happens to solve the task” as your behavior, and “first” here means “first one the algorithm is likely to discover”, aka: the “simplest” solution. hence my question: can you somehow measure this? ask a genetic algorithm to do some task and see whats the statistical distribution of its solutions and how fast it converges to those types of solutions, then enforce that it isnt allowed to generate that particular class of solutions and more to get the “second simplest” solution and repeat till it stops being able to solve the task
