Abstract
Development of Fuzzy Cognitive Maps (FCMs) that accurately describe a given dynamic system is a challenging task which in many cases cannot be fully completed based solely on human expertise. Some of the reasons behind this limitation include potential bias of the human experts and excessive size of the problem itself. However, due to the lack of automated or semi-automated methods that would replace or support designers, most of existing FCMs were developed using expert-based approaches. Interestingly, in the recent years we have witnessed the development of algorithms that support learning of FCMs from data. The learning corresponds to the construction of connection matrices based on historical data presented in the form of multivariate time series. Since the FCM may include feedback loops and they incorporate nontrivial transformation functions, forming these models from data is a complex task that requires searching through a large solution space. The existing automated learning methods are based either on the Hebbian learning or they apply evolutionary algorithms. This chapter formulates the task of learning FCMs and describes the corresponding design challenges. We present a comprehensive survey of the current expert-based and semi-automated/automated methods for learning FCMs. The leading learning methods are described and analyzed both analytically and experimentally with the help of a case study. We also contrast computational approaches versus expert-based methods and outline future research directions.