Research Abstracts - 2006

People generally use the Web to access passive information. Educational materials on the Web are generally fixed presentations of predetermined content. Intelligent books are different, because the information they present is only partly composed by the authors: most of an intelligent book is dynamically constructed and accumulated through conversations with its readers. More concretely, an intelligent book understands the principles and strategies of the subject matter, and uses those to answer questions, to fill in details, and to accumulate new examples as suggested by the conversations with its readers.

An intelligent book has several components:

1. Content presentation (including graphics and other media), as one would expect in an actual Web-based book, together with appropriate user interfaces.

2. The content presentation is backed up by a modelling system. In an intelligent book on circuit theory, for example, all the diagrams and equations would be active, with the equations linked to an algebraic manipulator and the diagrams linked to symbolic and numeric simulators. Thus, a student could point to any node on a circuit diagram and ask for the voltage at that node in symbolic form.

3. The modelling system is backed up by a simple reasoning and deductive system that uses constraint propagation and dependency-directed backtracking in order to make simple deductions and explain its deductions. For example, a student could point to a node on a circuit and ask the book to explain why the voltage is what it is; or point to an equation and ask the book to explain its derivation.

4. Integrated into the book are techniques for assessing a reader's progress and forming simple assessment. For example, the same underlying modelling and reasoning components that are used to explain the derivation of an equation, can be used to check a student's derivation of the equation, and offer hints to students who are stuck.

5. All of this is incorporated into a framework for automatically improving the books' presentation, based on "learning" from large numbers of users accessing the book via the Web. One aspect of this is the ability for the book to learn new examples, based on questions and suggestions from users. Another is increasing reliability of the assessment, based on aggregating the behaviour of large numbers of users.

   In our work so far, we have concentrated on examples from digital electronics and discrete mathematics at Cambridge, and transistor circuits at MIT. The MIT materials have been tested in experimental versions of the introductory electronic subject, which uses circuit models backed up by constraint propagation and truth maintenance systems to guide students through problems in amplifier design.

   In the figure below, a student is attempting to synthesize an amplifier, using a heuristic "crude bias model" to deduce value for the components. At the point shown in the figure, the student's proposed design violates the given specifications for the amplifier, and the system signals this as a contradiction.

   This work is funded by the Cambridge-MIT Institute.