Version Space Algebra and its Application to Programming by Demonstration

Machine learning research has been very successful at producing powerful, broadlyapplicable classification learners. However, many practical learning problems do not fit the classification framework well, and as a result the initial phase of suitably formulating the problem and incorporating the relevant domain knowledge can be very difficult and typically consumes the majority of the project time. Here we propose a framework to systematize and speed this process, based on the notion of version space algebra. We extend the notion of version spaces beyond concept learning, and propose that carefullytailored version spaces for complex applications can be built by composing simpler, restricted version spaces. We illustrate our approach with SMARTedit, a programming by demonstration application for repetitive text-editing that uses version space algebra to guide a search over text-editing action sequences. We demonstrate the system on a suite of repetitive text-editing problems and present experimental results showing its effectiveness in learning from a small number of examples.