Learning and forgetting in nonstationary environments : a multivariate change detection mechanism

In a contemporary world where exchanges of massive amounts of information have become the norm, it is useful to understand how human brain learns, collects relevant information and deletes the unnecessary one, in order to abstract learning mechanisms for systems working in nonstationary environments. At the same time, the pervasive use of technology, especially sensors, has become an extension of human brain and capabilities. Hence, the enormous quantity of data collected every day by these sensors can represent a very important source of information, and must be properly analysed, collected or forgotten. The abilities of learning and forgetting implemented by the human brain are necessary since the world we live in is governed by nonstationary, unpredictable rules. Hence detecting a change, a new arising situation, is fundamental. Starting from the structure of human memory, from its associativity and plasticity and from learning and forgetting processes, we derived an engineering-oriented perspective to propose intelligent solutions for embedded and cyber-physical systems. In particular a Change Detection Mechanism has been developed able to detect changes in multivariate (possibly very large) data streams. Differently from the majority of change detection tests existing in literature that require univariate distributions (1D) or the knowledge of the process before and after the change, our mechanism has been designed to work with multivariate, non parametric distributions, hence not requiring any a-priori information about the process or the change. Two approaches have been thought depending on whether we are interested in observing a change in the mean vector only and without requiring a training phase, or in both the mean vector and covariance matrix yet with the necessity of a training phase. Moreover, a substantial part of this work has focused on the estimation of the Average Run Length (ARL) and control limit values. Since it was not possible to evaluate them theoretically, simulations have been conducted over in-control data streams in order to collect the results. These values represent a central part of our CDTs because they drive the configuration of the entire algorithm. To conclude, the entire work has been successfully validated conducting experiments on both synthetic and real datasets.