Towards an Ensemble Framework for Assisting in Synthesis Tasks

It is a common practice when problem solving to seek the advice of others as to what decision to make next. One may seek advice from a single trusted colleague, but commonly one seeks advice frommultiple sources, weighing the advice from each before making a decision (Polikar 2006). This behavior is mimicked by ensemble methods for machine learning which combine outputs from multiple independent components to arrive at conclusions which, if the ensemble was constructed properly, can be more accurate and reliable than its individual components. Ensemble methods have enjoyed much recent attention from machine learning researchers. Classification systems have been studied extensively through ensemble methods, and though efforts have also been made to study clustering and regression ensembles, some have suggested wider application of ensemble methods (Rokach 2009). We seek to assist in synthesis tasks involving the design of structures, examples including plans (Kim and Blythe 2003; Aha, Breslow, and Munoz-Avila 2001) and workflows (Leake and Kendall-Morwick 2009). As such tasks can involve a high level of sophistication and complexity, it is not straightforward to apply machine learning techniques aimed towards analytical tasks (e.g., classification, regression) (Aha and Wettschereck 1997). To simplify the design task, incremental refinements are often sought involving explicit aspects of an incomplete or incorrect structure. These refinements can be presented as recommendations in a userdriven process where the AI system provides assistance to a human author. We simplify such recommendations to two core components, the problem and the solution, in order to reduce the problem of generating recommendations into two analytical tasks, described in detail in the following section. Most studies of ensemble methods have focused on varying the training set, feature set, or random values initializing a single inducer to produce a diverse set of classifiers. Even though there is no change in the inductive bias for each component of the ensemble, ensemble methods can increase the