GENERATING ADVICE BY MONITORING USER BEHAVIOUR

Abstract

Computer systems are becoming increasingly complex. Active assistance from the human-machine interface is required to exploit this growing sophistication fully. Traditional methods of providing assistance are inadequate in two ways. first, they are passive, requiring the human to detect (perhaps implicit) problems and assuming the human will make "appropriate queries". Secondly, they are static and so must explain everything in great detail. Active assistants are required to recognize what information is most important and make it prominent. This thesis explores a small part of this problem. A computer coach unobtrusively monitors user performance, attempts to recognize inefficient use of unawareness of important facilities, then suggests alternatives. Various design issues are discussed. The implementation of a prototype coach built for the Emacs text editor is described. Informal evaluation shows that the detection of unused concepts is a powerful yet easily implemented technique for generating helpful advice. The recognition of instances of pre-stored sub-optimal sequences of operations is shown to be less prolific and more difficult to implement, but to provide a greater degree of situation sensitivity.