9.9 Further Reading 397 user at each point. These rules are conditional so that users can teach classification tasks such as sorting files based on their type and assigning labels based on their size. They are learned incrementally: The agent adapts to individual users by record- ing their interaction history. Many difficulties arise. One is scarcity of data. Users are loath to demonstrate several iterations of a task--they think the agent should immediately catch on to what they are doing. Whereas a data miner would consider a 100-instance dataset miniscule, users bridle at the prospect of demonstrating a task even half a dozen times. A second difficulty is the plethora of attributes. The computer desktop envi- ronment has hundreds of features that any given action might depend on. This means that small datasets are overwhelmingly likely to contain attributes that are apparently highly predictive but nevertheless irrelevant, and specialized statistical tests are needed to compare alternative hypotheses. A third difficulty is that the iterative, improvement-driven development style that characterizes data mining applications fails. It is impossible in principle to create a fixed training and testing corpus for an interactive problem, such as programming by demonstration, because each improve- ment in the agent alters the test data by affecting how users react to it. A fourth difficulty is that existing application programs provide limited access to application