|
Abstract:
|
In this doctoral dissertation the modelling process has been taken into consideration in the presence of uncertainty. Two types of problems were analyzed: one is the optimization of the benefit/costs tradeoff during the distribution of the projects and the other is the classification of data described by the attributes among which some are missing. The basic problems during the modelling of the decision making in the presence of uncertainty are the choice of the adequate treatment of uncertainty and the choice of the method for making a decision.
One of the aims of the work is investigating the benefits of applying the metaheuristic algorithms on the considered optimization problems. The main measure for the evaluation of their performances is the value of objective function (for both problems: optimization of benefit/costs tradeoff during the project scheduling and clustering of incomplete data). Considering the project scheduling problem the level of satisfaction related to the problem constraints could also be taken into account. The other evaluation criteria of the applied metaheuristic methods is the time required for finding the solution. The influence of the parameters which control the algorithms of the metaheuristic methods is examined, as well as their appropriate values leading to the maximum performances of the implementation could be reached on the tested examples of the considered problems.
As for the optimization problem of the profit/costs tradeoff, the uncertainty is modelled by applying the triangle fuzzy problems and then the metaheuristic methods, simulated annealing and genetic algorithm were applied for solving the obtained fuzzy optimization problem. The tested problems are formulated by the fuzzification method which was suggested by (Ribeiro et al. 1999). The represented experimental results for the set of fuzzy problems show the efficiency of the applied methods: simulated annealing and genetic algorithm. Genetic algorithm seems to produce slightly better solution than the simulated annealing. However, both methods out performed the existing form the literature for about 20%.
The secund part of the work deals with the clustering data problem with the missing values of the attributes and making decisions in such circumstances. The main phases in solving the considered problem are finding the most appropriate distance, which will be used in the cases when the data are missing for some reasons and choosing the method for solving the clustering problem. As the theoretical and practical contribution, the metric, based on the logic principles, was proposed. By applying the probability, the theorem was proved defining the values of the weighting coefficients related to attributes that describe the objects for clustering. The proposed metric was implemented in the variable neighborhood search metaheuristic method as well as in some of its modifications. The implemented methods have been applied on the real life problems from the literature. Classifying the patients who suffer from some auto-immune diseases,
stored in the database of Clinical Centre of Serbia, the precision of the clustering of 93.33% was achieved. As another real life example, seven databases of the European Commision (Board), which contain the data for the mail service, have been analyzed. The clustering efficiency of 90% - 96.96% was achieved. In order to compare the efficiency of the approach based on the variable neighborhood search method, nine databases available on the internet were used and the obtained results were compared with the existing ones from the literature. The experiments showed large stability of variable neighborhood search method: in eight out of nine cases the best solution was reached in all hundred repetitions. Besides that, the quality of the obtained solutons have considerably surpassed the results from the literature. |