The objective of this paper is to apply fuzzy cognitive map (FCM)-related techniques to (1) extract causal knowledge from a specific problem-domain, (2) construct a hierarchical knowledge base, and (3) perform a bi-directional inference. The causal knowledge base built by FCM can be described as a matrix form, guaranteeing mathematically compact operation compared with a production (if–then) knowledge base. Based on the causal knowledge base, we can break down a given decision problem into a multi-leveled one. Then bi-directional (downward or upward) inference can be applied to the multi-leveled problem to find a more robust solution. We applied our method to a stock investment analysis problem which is typical of highly unstructured problems in OR/MS fields. © 1997 by John Wiley & Sons, Ltd.