Referring back to my 16 September 2010 blog defining several different sorts of intelligent systems and to the definition of intelligence in my letter in Skeptic, vol 12, #3, pg 14, 2006. A system having a performance element only (no evaluation function) has functions that map sensory inputs to motor outputs and yet it still maximizes some quantity (see Barrow and Tipler, The Anthropic Cosmological Principle, OUP, 1986, pg 151 and Yourgrau and Mandelstam, Variational Principles in Dynamics and Quantum Theory, Dover, 1968, pg 176). But this may not be doing a good job of maximizing the things we want, utility, reproductive success, or lifespan. Evolution acting on a set of such agents can tend to improve this over the time scale of generations. If the environment is not changing no further learning might ever be required.
Adding an evaluation function and some sensors that can measure reproduction, pain, damage, etc. can tend to improve the "utility" on a time scale shorter than an agent's lifespan. This is important, for instance, in dealing with environments that are rapidly changing.
We do not necessarily expect to ever be optimal we are simply seeking to improve over time.