For several centuries the so-called first principles models have dominated the natural sciences. However, for a number of practical engineering problems they are difficult or even impossible to build. Another alternative is to use so-called "black-box" models (polynomial, regression models, neural networks). They can fit the data with arbitrary precision, but they are not transparent enough: their coefficients and structure is not directly related to the system being modelled. Fuzzy rule-based models and especially Takagi-Sugeno (TS) fuzzy models have gained significant impetus due to their flexibility and computational efficiency. They have a quasi-linear nature and use the idea of approximation of a nonlinear system by a collection of fuzzily mixed local linear models. The TS fuzzy model is attractive because of its ability to approximate nonlinear dynamics, multiple operating modes and significant parameter and structure variations. On-line learning of TS fuzzy models involves recursive, non-iterative clustering responsible for model structure (rule base) learning and recursive consequent parameter estimation. eTS is based on the assumption that the model structure evolves gradually instead of being known a priori. It is important to note that this evolution is much slower than the evolution of the model parameters. For the eTS the notion of informative potential of the new data sample (accumulated spatial proximity measure) is very important. It has been first introduced in the mountain clustering approach and then refined in the subtractive clustering approach. It is used as a trigger to update the rule-base. It is a great advantage of this approach that the learning can start without a priori information and only a single data sample. This interesting feature makes the approach potentially very useful in autonomous, robotic, and smart adaptive systems. Recently the evolving Takagi-Sugeno (eTS) fuzzy models and the method for their on-line identification have been extended to the MIMO case. In the present paper, the algorithm is further simplified by replacing the Gaussian membership function with Caushy one (which has important implications for real-time realisations of the algorithm on chip), the Euclidean distance with Mahalonobis one. The problem of interpretability and simplification of the rule-base during the on-line learning is studied and an effective mechanism is proposed. The upper limit of the possible number of rules in the rule base is also defined. The radius of influence of each fuzzy rule is considered to be dependent on data and its orientation. Simulation results using a well-known benchmarks and with real data are presented.