definition: "A program that [...] provides assistance to a user dealing with a particular application. Such agents learn by watching over the shoulder of the user and detecting patterns and regularities..." (Maes)

This approach aims to overcome the limitations of the direct manipulation paradigm when a system becomes too complex to understand or manage.

(slide presentation by Radhakrishnan)

options: (Maes)

forces: several criteria will help to choose between the different options: example:

Mail-filtering [2]: uses RoteLearning and clustering techniques by storing a set of situation-action pairs. In a given situation, the proper action to trigger is calculated the following way:

  1. calculate the distance between the current situation and all the memorized situations
  2. choose the most appropriate action amongst the N closest situations (k-closest neighbors clustering technique)
  3. calculate the "score" of the action, which determines if the action should be only suggested ("tell-me" threshold) or triggered ("do-it" threshold)

It is possible to "forget" old cases, in order to help the system to adapt and to avoid it to slow down. Furthermore, there are 3 other types of learning that can help improve the system: other examples:

see also: the following borderline cases of interface agents:


See book.

back to SoftwareAgentCourse

(last edited November 4, 2008)
Find Page by browsing or searching