Information Technology Reference
In-Depth Information
contextual area. The decision of the most appropriate actions requires advanced Artificial
Intelligence Technique to satisfy a plethora of application domains in which interaction and
conclusive results are needed. This only is possible with Intelligent Systems equipped with
high processing speed, knowledge bases and an innovative model for designing
experiments, something will happen in this decade.
7. Conclusions
Many real world problems belong to a special class of problems called NP-hard, which
means that there are no known efficient algorithms to solve them exactly in the worst case.
The specialized literature offers a variety of heuristic algorithms, which have shown
satisfactory performance. However, despite the efforts of the scientific community in
developing new strategies, to date, there is no an algorithm that is the best for all possible
situations. The design of appropriate algorithms to specific conditions is often the best
option. In consequence, several approaches have emerged to deal with the algorithm
selection problem. We review hyper-heuristics and meta-learning; both related and
promising approaches.
Meta-learning, through machine learning methods like clustering and classification, is a
well-established approach of selecting algorithms, particularity to solve hard optimization
problems. Despite this, comparisons and evaluations of machine learning methods to build
algorithm selector is not a common practice. We compared three machine learning
techniques for algorithm selection on standard data sets. The experimental results revealed
in general, a high performance with respect to a random algorithm selector, but low perform
with respect to other classification tasks. We identified that the Self-Organising Neural
Network is a promising method for selection; it could reaches 100% of accuracy when
feedback was incorporated and the number of problem characteristics was the minimum.
On the other hand, hyper-heuristics offers a general framework to design algorithms that
ideally can select and generate heuristics adapted to a particular problem instance. We use
this approach to automatically select, among basic-heuristics, the most promising to adjust a
parameter control of an Ant Colony Optimization algorithm for routing messages. The
adaptive parameter tuning with hyper-heuristics is a recent open research.
In order to get a bigger picture of the algorithm performance we need to know them in
depth. However, most of the algorithmic performance studies have focused exclusively on
identifying sets of instances of different degrees of difficulty; in reducing the time needed
to resolve these cases and reduce the solution errors; in many cases following the strategy
"the -winner takes-all". Although these are important goals, most approaches have been
quite particular. In that sense, statistical methods and machine learning will be an
important element to build performance models for understanding the relationship
between the characteristics of optimization problems, the search space that defines the
behavior of algorithms that solve, and the final performance achieved by these
algorithms. We envision that the knowledge gained, in addition to supporting the growth
of the area, will be useful to automate the selection of algorithms and refine algorithms;
hiper-heuristics, hybridization, and meta-learning go in the same direction and can
complement each other.
