A method is proposed for expressing reduced member dimensions of an optimization problem. This method is based on mixture model for probabilistic principal component analyzers for feature extraction of given member section list, whose parameter can be determined using an expectation-maximization algorithm. In this paper, a new efficient method enables to find an optimum set of member sections of the steel building frames so that total weight would be minimum subject to constraints on interstory drifts under design spectrum-compatible earthquakes.