Conversation

I recall a nice optimization method called SPSA: successive perturbation stochastic approximation. Trick was to half-ass it in early steps, and refine it only when you get close to answer. No reason to get intermediate answers perfect. Excellent mediocrity engineering there.