Although Ga may be used for multiobjective optimization and it gives pareto optimal solutions in place of single solution but NN gives only a single solution at a time. Then which set is considered as ioptimum solution?
I think, you may find the optimum solution by trial and error method by using artificial neural network. Optimum solution may be decided based on objectives.
I'm not sure in multi-objective optimization, but I've already seen the application of NN in order to detect promising areas in the search space, considering a single evaluation function. The multi-objective case is more complicated, however I would say that you can used it in the same way as well.
I do not know how to apply one neural network for multiobjective optimzation problems. In these problems all Pareto optimal solutions are considered solutions, in the normal case there are infinitely many of them. An often adopted idea is to apply a scalarization like e.g. weighted sum, Zeleny's displaced ideal or Wierzbicki's reference point approaches. Using these (each of them is parametrized) produce one or several Pareto solutions and present these to a decision maker to evaluate them. The result will normally be the wish to change certain aspects of some of the solution - leading to a re-solve with changed parameters. Thus the decision maker learns about the properties of the model (if I want to improve one objective - what do I have to sacrifice wrt. to the others?)
Repeating this process might relsult in an accepted compromise solution among the Pareto optimal ones.
As search algorithms imply no guarantee for optimality, I doubt that GS or NN could ensure yielding Pareto optimal optimal points.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence