Learning systems such as multi-layer feed-forward neural networks, wavelet networks and so on need appropriate learning data (input data and teaching output data). These methods are not so useful in case when we cannot get the appropriate learning data.

Even in this case, it is not so difficult to evaluate the system output for arbitrarily applied input. The learning data of input-output pairs with their evaluations are easily obtained and thus is easily used for modeling the system. SOR (Self-organizing Relationship) network is a modeling tool, which can be established by a set of input-output data and corresponding evaluation. This SOR network can act as a knowledge acquisition system and also act as a fuzzy inference engine.

The linkage among the units in competitive layer is fixed and not flexible, and thus not used for complicated systems. In this plenary talk, the advanced learning process is presented for the original SOR network by employing evaluation-based TRN (topology representing network). By this learning, the linkage among the units in the competitive layer can be more flexible and thus used for modeling of much more complicated systems. The application of the SOR network established by this learning process to a manipulation control is also presented.