Learning complex goals with iterated amplification