MIC 2007: The Seventh Metaheuristics International Conference 0106-1 Parallel noising methods embedded in an adaptive memory Maxence Laurent ∗ , ´ Eric D. Taillard ∗ , Michel Toulouse † , Teo Crainic † ∗ HEIG-Vd, University of Applied Sciences of Western Switzerland Route de Cheseaux 1, Case Postale, CH-1401 Yverdon-les-Bains {Maxence.Laurent,Eric.Taillard} at heig-vd.ch † CIRRELT, Universit´ e de Montr´ eal C.P. 6128, Succursale Centre-Ville, Montreal H3C 3J7, Canada {Michel.Toulouse, theo} at crt.umontreal.ca This paper presents a high-level technique for parallelizing noising methods. To illustrate the technique, it is applied to the traveling salesman problem (TSP). It mixes 3 concepts proposed by various authors: 1) Noising methods proposed by Charon and Hudry [4]. The basic idea of noising methods is to add a random noise either to the problem data or to the objective function. 2) Adaptive memory methods of Taillard et al. [5, 6] that unify several metaheuristic concepts. Memetic algorithms, ant colonies hybridized with a local search, scatter search, vocabulary building and path relinking methods are analyzed with an adaptive memory point of view in [3]. So far, noising methods have not been embedded in the adaptive memory frame. 3) Guided cooperative search of Le Bouthillier, Crainic and Kropf [1, 2] who propose a special mechanism for exploiting an adaptive memory: The idea of [1, 2] is to try to get more pertinent information from the solutions stored in the memory, and especially from the bad ones. Mixing these 3 techniques together has been done as follows: Memory: The memory (or data warehouse) is constituted of solutions. Each solution is classed into 3 categories : elite, intermediate and bad solutions. Each solution contains several components (for the TSP, a component is an edge: a route from one city to the next one). So, each component can be classified into 0, 1 or several of the 3 classes (no solution of the memory contains a given component, a component belongs to solutions of a single class or a component belongs to solutions of several classes). Memory initialization: The memory is initialized with solutions created with the the Quick-Boruvka procedure, as implemented in the Concorde software. These solutions are improved with the Chained Lin-Kernighan (CLK) procedure implemented in the Concorde software. Noising: Before launching CLK, the length of each edge is perturbed by a value that depends on a value 1 >r> 0. This value r linearly decreases with the iteration number. For perturbing the length of the edges, we first multiply this length by a factor uniformly distributed between 1 - r and 1 + r. If the edge only belongs to elite solutions, the perturbed length is then diminished by a factor 1 - r. If the edge only belongs to bad solutions, the perturbed length is then increased by a factor 1 + r. Building a new solution: The quickest way to build a new solution is to start from a solution already in memory. So, we took the best solution stored in memory. Since the length of Montreal, Canada, June 25–29, 2007