training skills that are slightly suboptimal but generalize better is not irrational if you estimate a high likelihood that the game will change faster than you can retrain to a new narrow winning skill