Learning is optimized when we fail 15% of the time