In this paper, a hybrid multi-concept acquisition system HMCAS is proposed. HMCAS can perform incremental supervised learning on arbitrary sequences composed of analog or binary inputs. The kernel algorithm of HMCAS, named HMCAP, which integrates symbolic and neural learning based on the probability of instance space, has the ability of generating concept descriptions in the form of hybrid decision tree. The prototype system of HMCAS has been applied to the field of typhoon forecasting and achieved successful result.