This dissertation presents the proposal and the development of a totally automatic data mining system. The main objective is to create a system that is capable of extracting obscure information from complex databases, without demanding the presence of a technical specialist to configure it. The Hierarchical Neuro-Fuzzy Binary Space Partitioning model (NFHB) has produced excellent results in pattern classification and time series forecasting tasks. Additionally, it provides important features that are not present in other similar systems, such as: automatic learning of its structure; ability to deal with a larger number of input variables, thus increasing the range of possible applications; and generation of linguistic rules as a result of its training process. However, this model depends on a complex configuration process before the training is performed, hindering to achieve a totally automatic system. The model proposed in this Dissertation tries to optimize the NFHB system parameters by using the genetic coevolution technique, thus creating a new automatic data mining system. This work consisted of four main parts: evaluation of existing systems used in data mining; study of the NFHB system and definition of its main parameters; development of the automatic hybrid neuro-fuzzy-genetic system for data mining; and case studies. In the study of existing data mining systems, the aim was to find a suitable model that could yield good results and still be automated. Several techniques have been studied, among them: Statistical methods, Decision Trees, Rules Association, Genetic Algorithms, Artificial Neural Networks, Fuzzy and Neuro- Fuzzy Systems. The NFHB System was chosen for inference and rule extraction in the data mining process. In this way, this model was carefully studied and its most important parameters were determined. Moreover, input variable selection techniques were investigated, to be used with the proposed model. Finally, a set of parameters was defined, which must be determined automatically for the complete system configuration. A hierarchical coevolutive genetic model was created to execute the system optimization task with efficiency. Therefore, a hierarchical architecture of genetic algorithms (GAs) was created, where the GAs execute complementary optimization tasks. In this stage, the best genetic operators, the GAs configuration, the chromossomes representation, and evaluation functions were also determined. The best set of parameters found was used in the NFHB configuration, making the process entirely automatic. In the case studies, various tests were performed with benchmark databases. For forecasting problems, six electric load series were used: Cerj, Copel, Eletropaulo, Cemig, Furnas and Light. In the pattern classification area, some well known databases were used, namely Glass Data, Wine Data, Bupa Liver Disorders and Pima Indian Diabetes. After the tests were carried out, a comparison was made with known models and with the original NFHB System, configured by a specialist. The tests have demonstrated that the proposed model generates satisfactory results, producing, with an automatic process, similar errors to the ones obtained with a specialist configuration, and, in some cases, even better results can be obtained. Therefore, a user without any technical knowledge of the system, can use it to perform data mining, extracting information and knowledge that can help him/her in decision taking processes, which is the final objective of a Knowledge Data Discovery process.