In 1975, Holland publishes Adaptation in Natural and Artificial Systems. His genetic algorithms converge toward configurations he had not programmed. Random mutations and artificial selection produce solutions no one had designed.
The solution space exists before the first calculation. Each optimization problem defines a fitness landscape, a geometry where each point corresponds to a possible solution and each altitude to its performance. The algorithm does not create this landscape. It explores it. The optimization peaks were already there, carved into the problem's constraints.
The optimal future constrains each iteration from the beginning. Each algorithm generation carries the trace of the final result, the mutations that survive are those pointing toward the optimum, even if the algorithm does not yet know it. Convergence reveals a preexisting geometry.
Selective breeding exposes a deeper asymmetry: selection cannot overflow the possible. The breeder defines fitness criteria, but this decision determines the landscape he has not yet examined. To formulate what one seeks is already to constrain what one can find. The creative act is not in selection. It is in choosing the problem.
Doctrine
What one seeks determines what one can find. What one can find existed before one began seeking. Invention is not in the algorithm, it is in problem formulation.
Vecteur ouvert
If each optimization problem defines a preexisting fitness landscape, then two problems whose constraints converge toward the same solution space share a geometry, independent of the domain in which they are formulated. Have aeronautical engineering and bird wing morphology solved the same problem, or have they discovered the same solution in two distinct landscapes?
