A rule-based expert system has been built for selection of both retaining wall types and groundwater control methods in deep excavations in Wuhan city. For this expert system, a new type of generation rule is developed in which one condition is able to be defined with a "third state" that not only contributes directly to reaching the conclusion in a rule, but also factors into calculating the reliability of the conclusion. The traditional backward chaining technique has been improved to accommodate the change of a rule type and a fuzzy backward chaining method IRO has been established to increase reasoning flexibility. Using IRO as a fundamental element, it is convenient to form a complicated reasoning network in the inference engine. Finally, two knowledge bases are built from more than 100 case histories and other resources, and the new expert system proves to be effective in case studies. (C) 2003 Elsevier Ltd. All rights reserved.