**2. Simple and interpretable hydrological models using genetic programming**

The major drawback of all the data driven modelling approaches is the black box nature of these models, i.e., the user cannot easily identify what is happening in model which computes the outputs corresponding to the inputs supplied to the model. One of the key advantages of genetic programming as a modelling tool is its ability to develop simple hydrological models. The simplicity of the models is close associated with their interpretability. The simpler the models are the better they can be interpreted. This in turn helps in assessing the contributions of different members of the predictor set or inputs in making a particular prediction. Selle and Muttil (2011) utilized this capability of GP to test the structure of hydrological models to predict deep percolation response in surface irrigated pastures. Data obtained using lysimeter experiments were used to develop simple models using genetic programming. The developed models were simple and interpretable which helped in identifying the dominant processes involved in the deep percolation process. Often the developed models could be expressed as simple algebraic equations. The dominant processes identified compared well with the same as used in conceptual models. The study also investigated the recurrence of the models developed using GP in multiple runs and found out that they were consistently coming up with the same model for a given level of complexity of the model. However, the study also reported that as the level of complexity increases recurrence of the generated model were affected and the physical interpretability of the models decreases and hence careful understanding of the complexity of the system is to be considered before a level of complexity is chosen for the GP models.

This however, illustrates that carefully developed GP models remain mathematically simple and are readily interpretable to the extent that the dominant processes which influence the prediction could be readily identified from the model structure. When carefully implemented models can throw light into and identify the key physical processes contributing to the phenomenon predicted and hence the development of the model. This is an important feature lacking from many of the data mining based prediction models resulting from which these modelling approaches are often earmarked as "black-box" models. "Black-box" nature of the prediction models often result in the limited use of such models for practical predictive applications.
