The LASCA strategy

The project LASCA aimed at testing a novel strategy to address large scale optimization problems, when using metaheuristics as a solver. It is known that the meta-heuristics behaviour (such as in evolutionary algorithms) degrades quite considerably with the dimensional growth of the search space.

The conjecture behind the LASCA approach is that a reduction in the search space dimension, even at the cost of some loss in precision, should lead, at least in some classes of problems, to a speed up in convergence of some meta-heuristic and, in a favourable scenario, to a convergence with higher precision (for instance, from escaping to getting trapped in local optima).

The strategy put to test was the following:

1.     Given a problem, run a meta-heuristic optimization algorithm for some generations, and collect data about the progress of the search.

2.     Use these data to train an autoencoder

3.     Use the autoencoder to obtain a projection of the search into a smaller dimension space

4.     Organize the progress of the search for the optimum in the new space

5.     At some point, return to the higher dimension space and tune up the solution.