Instructible agents

Abstract

This dissertation explores the design of inference and interaction methods that would enable end users to teach a computer the characteristic, discriminating features of a set of data. The envisaged application is an agent that learns to select and edit data, or to observe and maintain relationships between data. Users would teach by giving examples, hints and partial specifications. The work described here develops design requirements, implements inference and interaction techniques, and gathers empirical data on their usability. The research method interleaves analysis, implementation and user testing. Chapter 1 motivates the research with a detailed worked example. It defines design goals and functional requirements, on which previous work in machine learning and programming by demonstration is assessed, revealing that no system enables users to teach concepts in a rich representation by giving both examples and ambiguous hints. Chapter 2 presents empirical data on methods of instruction which users readily adopt to teach a computer agent. A preliminary, underspecified design is simulated manually. A diverse group of users consistently develop similar sets of commands, and learn the agent's language from its verbal feedback. Chapter 3 explains the use of dynamic bias to reduce the complexity of concept learning, and shows how instructions from a teacher (the user) can direct the bias. The result is a formal model of instruction based on classifying examples, hints and rules. Chapter 4 describes interaction techniques for teaching and controlling an agent, suitable for use in direct manipulation and menu-based interfaces. These techniques are iteratively designed and user tested in prototypes ranging from slideshows to partial implementations. Chapter 5 implements the first concept learning system whose dynamic bias enables it to learn from examples, ambiguous hints and partial specifications. Using multiple heuristics, it can choose the most justified interpretation of hints, and find plausible alternatives in case instructions are erroneous. When evaluated on tasks from the study in Chapter 2, the implemented system rivals the simulated agent's ability to learn from examples and hints.