The transformation from precipitation over a river basin to river streamflow is the result of many interacting processes which manifest themselves at various scales of time and space. The resulting complexity of hydrological systems, and the difficulty to properly and quantitatively express the information that is available about them, determine the challenge of Rainfall-Runoff (R-R) modeling. Accurate and reliable R-R models, however, are important because they can be used for scientific hypothesis testing, or for making prediction that can improve the quality or effectiveness of decisions related to water management issues. In recent years, Computational Intelligence (CI) has emerged as a promising field of research. It has increasingly found application in R-R modeling (see literature review in Chapter 2), and in the research community there exists a clear and urgent need to further investigate the application of CI techniques. The main objective of this research, therefore, was to use CI techniques in catchment-scale R-R modeling in order to find improved methods of developing and evaluating such models. Three fields of application are explored for these purposes: system identification, parameter estimation and data mining. In Chapter 3, a R-R model based on a well-known CI technique, an Artificial Neural Network (ANN) is developed. Some important issues regarding the development, calibration and performance of such models are highlighted and discussed. Chapter 4 deals with the application of evolutionary, multi-criteria algorithms to the calibration of ANN R-R models, along with a comparison with traditional single-criterion algorithms. A multi-criteria comparison of an ANN model and a conceptual hydrological model is subsequently presented in Chapter 5. In Chapter 6, a temporal clustering approach was employed to identify periods of hydrological similarity. The results were used to shown how the evaluation of a conceptual model can be improved to be more diagnostic in nature and how subsequent improvements to the model structure can be inferred. The following summarizes the investigations on each of the three applications. 1. Artificial Neural Networks as Data-Driven Models In Chapters 3, 4 and 5, ANNs have been used as data-driven R-R models to explore if CI techniques can simulate the R-R transformation adequately and how well they compare to conceptual hydrological models. The results show good model performance, but also show that the application of ANNs is not without problems, since they are quite sensitive to several subjective modeling choices. For example, the choice of model input, structure or training algorithm has a big influence in the accuracy and parameter uncertainty of the model (Chapters 3 and 4). Moreover, ANNs sometimes appear to be sensitive to timing errors (Chapter 3). With respect to conceptual models, ANNs show to be slightly better for short-term forecasting but their performance decreases with increased forecast lead times (see Chapter 5). All in all, the development, calibration and evaluation of ANN R-R models, and the underlying uncertainties involved, which are subject to ANN model structure, objective functions, optimization algorithms, initialization, etc., continues to be a complex and opaque field. More insight in these issues is needed before the use of data-driven techniques such as ANNs can either be recommended or discouraged as adequate alternatives to traditional R-R models. 2. Computationally Intelligent Parameter Estimation CI parameter estimation algorithms have been applied to calibration of both CI and conceptual models to test whether more information can be extracted from hydrological data and used to make better R-R models (see Chapters 4 and 5). Throughout this study, the sensitivity of results on the method of optimization are shown to be large. Whether it be in differences between various local algorithms (Chapter 3), between local and global algorithms, or between single-criterion and multi-criteria algorithms (both in Chapter 4), the choice of algorithm turns out to have large effects on model accuracy and uncertainty. MC algorithms such as the NSGA-II and MOSCEM-UA prove to be very valuable in R-R model calibration since they exploit the information in model and data in a better way than , making the models not only accurate but a lot more reliable (see Chapters 4 and 5). Their usefulness naturally depends on the choice of objective functions. In this work, a new objective function (the Mean Squared Derivative Error) was proposed that penalizes a model for errors regarding hydrograph timing and shape. It was shown to evaluate models in a uniquely different way compared to traditional objective functions. Finally, the powerful self-adaptive Differential Evolution algorithm was employed in Chapter 6 and showed to be effective in a model calibration procedure. Generally, it was concluded that CI parameter estimation methods are more effective compared to traditional techniques. 3. Hydrological Clustering In order to find and make use of dynamical patterns in hydrological data that are commonly ignored in model evaluation, data mining has been performed in Chapter 6. A temporal clustering approach based on the simple k-means clustering algorithm was successfully devised to partition the historical data into several periods of hydrological similarity. The parameter variability between the hydrologically similar periods was subsequently used to make diagnostic inferences leading to improvements in the proposed model structures. This diagnostic step represents a successful novel application of clustering techniques that addresses the challenging and fundamental hydrological issue of how to achieve improvements on the working model hypothesis. The overall conclusion of this work is that applications of CI in R-R modeling show a lot of promise. CI techniques can be considered powerful thanks to, for example, their general effectiveness in dealing with nonlinearity (e.g., ANNs as R-R models) and high-dimensionality (e.g., effectiveness of CI parameter estimation). Moreover, CI techniques provide an alternative viewpoint on hydrological data and models that is strikingly different from traditional techniques, and as such is able to extract information hitherto overlooked. Nevertheless, some pitfalls of CI techniques presented themselves. Therefore, there are advantages to approaches in which models use the characteristics of both knowledge-driven and data-driven modeling. This work shows that by correctly combining both process knowledge, modeling experience and intuition, it is possible to forge new model development and evaluation methods that combine the best of both worlds. This work has shown that even without the use of additional sources of information or observations from the field, current model evaluation practices can be significantly improved through careful scrutiny of data and model functioning with CI techniques. Two examples from this work include the MC approach presented in Chapters 4 and 5, and the diagnostic evaluation approach of Chapter 6. Both these approaches are meant to extract information from data and model that is commonly ignored when merely evaluating the difference between time series of model output and observations using a single statistic. New methods of comparing signatures of model and data such as these are valuable because they signify an improved ability to judge the quality of hypotheses about the real-world hydrological system on which the model is based. This ability is considered improved not only because it is more accurate and reliable but ultimately a diagnostic tool through which one can find how hypotheses can be improved.