Learning methods for dealing with complex nodes in Fuzzy Cognitive Maps. (R-11837)

Fuzzy Cognitive Maps (FCM) (Kosko, 1986) have been used in various areas as a simulation model for classification and regression. Recently, the literature shows an increase of studies concerning the learning of maps, reflected in several articles published by researchers from Hasselt University among others. The ability of FCMs to classify allows them to use this parameter to introduce reinforced learning techniques that support and improve prediction. FCMs are sometimes limited by the nature of real problems, some of which are difficult to represent. This proposal is based on the application of reinforcement learning and the introduction into the FCM of a new concept to model a network composed of networks. Therefore, the proposal is summarized in 3 main questions: 1) How to improve FCM learning by using reinforcement learning to update the weight matrix? 2) What actions should be taken on the relationships between nodes to obtain a greater reward to improve the performance of the map? 3) How to represent a node, so that it behaves like a map, to update its relations and weights in a different way depending on whether the relations are towards the inside in the node or towards the outside with an external node? The final goal of this investigation is to elaborate a map that overcomes these challenges to support the decision making.

Keywords:decision support systems, Fuzzy computation, Machine learning

Disciplines:Knowledge representation and reasoning, Machine learning and decision making, Neural, evolutionary and fuzzy computation