Multi-step environment learning classifier systems applied to hyper-heuristics

• Authors: Javier G. Marín-Blázquez;Sonia Schulenburg
• Venue: Proceedings of the 8th annual conference on Genetic and evolutionary computation
• Year: 2006

Heuristic Algorithms (HA) are very widely used to tackle practical problems in operations research. They are simple, easy to understand and inspire confidence. Many of these HAs are good for some problem instances while very poor for other cases. While Meta-Heuristics try to find which is the best heuristic and/or parameters to apply for a given problem instance Hyper-Heuristics (HH) try to combine several heuristics in the same solution searching process, switching among them whenever the circumstances vary. Besides, instead to solve a single problem instance it tries to find a general algorithm to apply to whole families of problems. HH use evolutionary methods to search for such a problem-solving algorithm and, once produced, to apply it to any new problem instance desired. Learning Classifier Systems (LCS), and in particular XCS, represents an elegant and simple way to try to fabricate such a composite algorithm. This represents a different kind of problem to those already studied by the LCS community. Previous work, using single step environments, already showed the usefulness of the approach. This paper goes further and studies the novel use of multi-step environments for HH and an alternate way to consider states to see if chains of actions can be learnt. A non-trivial, NP-hard family of problems, the Bin Packing one, is used as benchmark for the procedure. Results of the approach are very encouraging, showing outperformance over all HAs used individually and over previously reported work by the authors, including non-LCS (a GA based approach used for the same BP set of problems) and LCS (using single step environments).