Novel developments in computational intelligence with applications to data stream mining

点击翻译按钮获取中文摘要

Complex fuzzy logic is a recent generalization of the traditional fuzzy logic. Complex fuzzy truth values are vectors from the complex plane with a magnitude less than or equal to 1. There has been limited progress in elucidating the properties of complex fuzzy logic, and a few machine-learning approaches based on it have appeared (with our work prominent in both strands of research). However, it is believed that complex fuzzy logic is actually an infinite family of multivalued logics; and that the design space for learning algorithms based on it is correspondingly vast. Plainly, there is still an enormous amount of work to be done in understanding this area, and a clear opportunity for our research to make a lasting mark on the field. Our proposed program of research for the coming five years will continue our theoretical and applied research in this field, tackling key open questions such as: what are the classes of operators that form complex fuzzy conjunctions, disjunctions, and implications? What are their properties? What functions form useful complex fuzzy sets? How do we linguistically interpret complex fuzzy sets? How are complex fuzzy sets and logic most usefully realized in machine-learning algorithms, and what classes of problems are they best-suited to solve? To focus our program of research, we will pursue one class of applications. Based on current research results, we believe that complex fuzzy logic can be effective in mining data streams. Our ANCFIS learning architecture was very accurate in time-series forecasting (an instance of stream data), and we now seek to generalize this result. Accomplishing this goal will require us to answer the theoretical questions we have raised above, as they all directly relate to data stream mining. Our learning algorithms (using the identified operators) will extract synopses of the data stream in the form of complex fuzzy rules, which must then be interpreted to be actionable. Our initial stream mining problems are a pair of sensor-data applications: air-quality monitoring in Alberta's Wood Buffalo region (site of the oilsands), and livestock disease surveillance based on animal-mounted sensor platforms; both are of significant economic importance.