Algoritmos exatos para problema da clique máxima ponderada

Abstract

In this work, we present three new exact algorithms for the maximum weight clique problem. The three algorithms depend on an initial ordering of the vertices. Two ordering are considered, as a function of the weights of the vertices or the weights of the neighborhoods of the vertices. This leads to two versions of each algorithm. The first onde, called BITCLIQUE, is a combinatorial Branch & Bound algorithm. It effectively combines adaptations of several ideas already successfully employed to solve the problem, such as the use of a weighted integer coloring heuristic for pruning and branching, and the use of bitmap for simplifying the operations on the graph. The proposed algorithm outperforms state-of-the-art Branch & Bound algorithms in most instances of the considered in terms of the number of enumerated subproblems as well in terms of computational time. The second one is a Russian Dolls, called BITRDS, which incorporates the pruning and branching strategies based on weighted coloring. Computational tests show that BITRDS reduces both the number of enumerated subproblems and execution time when compared to the previous state-of-art Russian Dolls algorithm for the problem in random graph instances with density above 50%. As graph density increases, this difference increases. Besides, BITRDS is competitive with BITCLIQUE with better performance in random graph instances with density between 50% and 80%. Finally, we present a cooperation between the Russian Dolls method and the Resolution Search method. The proposed algorithm, called BITBR, uses both the weighted coloring and upper bounds given by the dolls to find a nogood. The hybrid algorithm reduces the number of coloring heuristic calls, reaching up to 1 order of magnitude when compared with BITRDS. However, this reduction decreases the execution time only in a few instances. Several computational experiments are carried out with the proposed and state-of-the-art algorithms. Computational results are reported for each algorithm using the main instances available in the literature. Finally, future directions of research are discussed.