In combinatorial optimization problems, finite collections of objects that meet specific criteria are studied in order to determine whether there is any optimal object. In most cases, although the search domain is finite, it can be of exponential size. Nowadays, it is possible to solve many combinatorial problems presented in real life using techniques based on integer programming.  However, combinatorial problems are extremely difficult in many cases and only near-optimal solutions are possible.   For these occasions, research efforts have focused on the application of meta-heuristic techniques. This is the case of this work, which is focused on solving complex high-dimensional combinatorial problems. In this kind of problems, exploring all possibilities to find the optimum is intractable, either for economic reasons (i.e. testing each combination is very expensive) or for computational reasons (i.e. being temporarilly intractable).

Specifically, this thesis proposes a domain-independent search architecture, which is  able to tackle large combinatorial problems, even in situations where there is little starting data. This architecture is based on Soft Computing techniques, combining a genetic algorithm based on real coding with artificial neural network-based models (multilayer perceptrons).  Thus, the genetic algorithm makes use of these perceptrons for fitness evaluations, when necessary.  The obtained system offers the required flexibility and versatility to be able to tackle with whatever combinatorial problem. 

Actual Soft Computing techniques have been reviewed in order to select the most appropriate to the pursued architecture. As a result, it was found that these techniques offer low cost, robust and flexible solutions. Furthermore, when acting in combination, they enhance their strengths while minimizing their disadvantages.

Furthermore, the developed Soft Computing architecture was applied to solve combinatorial problems of interest, both in the area of Combinatorial Catalysis and in the domain of Recommender Systems. Firstly,  requirements and needs of the problems to be solved within the scope of both domains were considered. Secondly, the proposed technique was used in the field of catalysis in order to optimize the conditions for different reactions, as well as to determine the best catalyst compositions suitable for reactions of different nature and complexity. Also, the proposed search architecture was applied to an entertainment domain in the field of Recommender Systems: the evaluation of films. For this study,  the MovieLens dataset was employed, which is a well-known benchmark in this field. Thus, the architecture was used to determine the preferences of certain users from the information available about them or from the information available about others users with similar preferences. 

Finally, the proposed architecture has been employed as a framework to obtain the  SoftCombi package, whose tools help in the intelligent design of experiments in the field of Combinatorial Catalysis.