Abstract. This paper describes an integrated system called KOSI (Knowledge Oriented Statistic Inference) for discovering functional relations from databases. Tile key feature of KOSI is that AI techniques and statistical methods are cooperatively used in the discovery process based on incipient hypothesis generation and evaluation, in which multi-search is performed in meta control. Two types of search, which use respectively different type of heuristics, are used in attribute calculation which is a kind of operation in which new attribute is generated as a function of the existing attributes. Third type of search, which is based on regression analysis, is mainly used for evaluating/selecting the best functional relation from the results of attribute calculation. Furthermore, a model-base and metaldomain knowledge are used for controlling the multi-search, and the methods of forming scopes/clusters can be used as a step of pre-processing before the search. KOSI is organized into multi-level structure for knowledge discovery and management efficiently. We try to provide a systematic manner of discovering functional relations, to support qualitative/quantitative discovery, and develop a more robust, general-purpose discovery system.