Programming by example: The human face of AI

It is argued that “human-centredness” will be an important characteristic of systems that learn tasks from human users, as the difficulties in inductive inference rule out learning without human assistance. The aim of “programming by example” is to create systems that learn how to perform tasks from their human users by being shown examples of what is to be done. Just as the user creates a learning environment for the system, so the system provides a teaching opportunity for the user, and emphasis is placed as much on facilitating successful teaching as on incorporating techniques of machine learning. If systems can “learn” repetitive tasks, their users will have the power to decide for themselves which parts of their jobs should be automated, and teach the system how to do them — reducing their dependence on intermediaries such as system designers and programmers. This paper presents principles for programming by example derived from experience in creating four prototype learners: for technical drawing, text editing, office tasks, and robot assembly. A teaching metaphor (a) enables the user to demonstrate a task by performing it manually, (b) helps to explain the learner's limited capabilities in terms of a persona, and (c) allows users to attribute intentionality. Tasks are represented procedurally, and augmented with constraints. Suitable mechanisms for attention focusing are necessary in order to control inductive search. Hidden features of a task should be made explicit so that the learner need not embark on the huge search entailed by hypothesizing missing steps.     

An experiment in the application of similarity-based learning to programming by example

Programming by example is a powerful way of bestowing on nonprogrammers the ability to communicate tasks to a computer. When creating procedures from examples it is necessary to be able to infer the existence of variables, conditional branches, and loops. This article explores the role of empirical or “similarity-based” learning in this process. For a concrete example of a procedure induction system, we use an existing scheme called METAMOUSE which allows graphical procedures to be specified from examples of their execution. A procedure is induced from the first example, and can be generalized in accordance with examples encountered later on. We describe how the system can be enhanced with Mitchell's candidate elimination algorithm, one of the simplest empirical learning techniques, to improve its ability to recognize constraints in a comprehensive and flexible manner. Procedure induction is, no doubt, a very complex task. This work revealed usefulness and effectiveness of empirical learning in procedure induction, although it cannot be a complete substitute for specific preprogrammed, domain knowledge in situations where this is readily available. However, in domains such as graphical editing, where knowledge is incomplete and/or incorrect, the best way to pursue may prove to be a combination of similarity- and explanation-based learning. © 1994 John Wiley & Sons, Inc.     

Cima: An interactive concept learning system for end-user applications

Personalizable software agents will learn new tasks from their users. In many cases the most appropriate way for users to teach is to demonstrate examples. Learning complex concepts from examples alone is hard, but agents can exploit other forms of instruction that users might give, ranging from yes/no responses to ambiguous, incomplete hints. Agents can also exploit background knowledge customized for applications such as drawing, word processing, and form filling. The Cima system learns generalized rules for classifying, generating, and modifying data, given examples, hints, and background knowledge. It copes with the ambiguity of user instructions by combining evidence from these sources. A dynamic bias manager generates candidate features (attribute values, functions, or relations) from which the learning algorithm selects relevant ones and forms appropriate rules. When tested on dialogs observed in a prior user study on a simulated interface agent, the system achieved 95% of the learning efficiency observed in that study.