Experience with a learning personal assistant
Communications of the ACM
The visualization of linear algebra algorithms in apt apprentice
ICCS'05 Proceedings of the 5th international conference on Computational Science - Volume Part I
| Hi-index | 0.02 |
Many users of workstation and PC tools often have to perform the same task again and again. For example, a secretary might have to send out a dozen email messages until she finds a free meeting room. Or someone preparing business charts has to draw many special tables with shadowing bars around. Unfortunately, today''s macro facilities of such tools do not support the end user enough in constructing the required automation functions. In this report we propose a mechanism, called dialog-based learning (DBL), that shall provide the user of software tools exactly with a mechanism to teach new functions or to give hints or additional information to a program on how to perform a task better. Two applications will be considered: The first one is our experimental system RAP, a room reservation apprentice that will eventually overtake a secretary''s task to search for a free meeting or lecture room. RAP analyzes the outgoing and incoming email and constructs a finite state machine that can repeat the task of asking all room administrators until a free room is found. The key of RAP''s learning is to ask the user for unknown message types (e.g., request, positive answer, etc.) and keyphrases (e.g., "need a room") and to collect them in a thesaurus. Our second application is a demonstrational graphics editor that allows the user to teach it new functions by giving a few examples. The graphics editor will sometimes ask the user for explanations by showing him a list of its geometrical hypotheses for a certain situation, e.g., "line l1 was doubled" or line "l1 touches line l2 in the middle" (the second hypothesis serves as a strong indication that "touching" is the intended property). By clicking at one of them the use tells the graphics editor his intention and helps it to construct a function and a menu item for a new function. Involving the user through dialogs is an easy way to support the heuristics for function induction from examples when the hypothesis space grows. Our claim is that -- with DBL -- a Programming by Demonstration system can synthesize new functions faster, that it can synthesize more complex functions than without DBL, and that the resulting functions will meet the user''s intentions better (because he or she took part in their derivation). Therefore one can see DBL as a kind of simple inference mechanism, but with a powerful outcome.