Title: DISCOVERY OF PROBABILISTIC RULES FOR PREDICTION
Keith C. C. Chan, Andrew K. C. Wong · 1989
Given an ordered sequence of objects (numbers, letters, symbols, observations, events, etc.) which is described by one or more attributes and is generated either deterministically or probabilistically. Suppose that the characteristics of the objects in the sequence is determined, to a certain extent, by the previous objects, the purpose of this paper is to present an inductive learning algorithm to analyze the inherent patterns in the sequence and to predict future objects based on these patterns. This inductive learning algorithm is divided into three phases: 1) detec- tion of underlying patterns in a sequences of objects; 2) construction of rules, based on the detected patterns, that describes the generation process of the sequence; and 3) use of these rules to predict the characteristics of the future objects. The learning algorithm has been implemented in a program known as the OBSERVER and it has been tested with both simulated and real-life data. The experimental results show that the OBSERVER is capable of discovering hidden patterns and explaining the behavior of certain sequence generating processes that a user is not immediately aware of or fully understood. For this reason, the OBSERVER may be employed to solve complex real world problems where predictions have to be made in the presence of uncertainty. It can also be employed, like some inductive learning programs, to be an aid for the knowledge acquisition process in the construction of knowledge-based systems.