``[E]volution should be free to explore the possibilities without the burden of human ``guidance''.''Alastair Channon
Evolutionary algorithms allow us to evolve a system to achieve a selected goal, without explicitly stating how to reach it. Specifying what to evolve towards, indirectly by specifying the fitness function, is an example of supervised learning. The fundamental problem of supervised learning used in an attempt to evolve complex behaviors, is that it is limited by the insight and creativity of the supervisor.
The situation is similar to programming a system to react intelligently. But then it is actually the programmer's intelligence that the system exhibits, and not its own. The extent of intelligent behavior is limited by the intelligence of the programmer. Many case based reasoning systems and knowledge-bases have been created, however most can only answer very narrow and specific questions.
As we have learned in section 3.3.2 (page ), intelligence has a general aspect to it. And when it is entirely pre-programmed, it is more aptly called an instinctive than intelligent. So far, we have been unable to exactly specify what intelligence is. In order to achieve emergence4.4 of intelligence, Alastair Channon argues for the withdrawal of the traditional `fitness function' based Genetic Algorithm (figure 4.1a). We are unable to determine such a function that produces intelligence. In `The Evolutionary Emergence route to Artificial Intelligence' [Cha96] Channon outlines evolution of a virtual world in which co-existence, interactions of species, acts as the selective force (figure 4.1b).
His view is shared with some others, including Thomas Ray. Succinctly put, this is what happens in the Tierran world: ``[O]nce the memory is filled with creatures, the creatures themselves become a prominent feature of the environment. Now evolution also discovers ways for creatures to exploit one another, and to defend against such exploitation.'' [Ray94, p. 245]
Channon concludes that the `Selection' principle of Darwin's theory is often misinterpreted and overemphasized. Reminding us that the theory is one of local change and adaptation, not of optimisation along an absolute scale of fitness. `Selection' is a mere abstraction of probability of inheritance of any property.
Erik de Bruijn 2007-10-19