Information Technology Reference
In-Depth Information
3.5.6.
Design of agent behaviors
In this phase, for each agent model we have refined and specified the agent
behaviors. For example, for the profile management agent we have applied our user
model put forward in [ANL 06a].
This model is made up of three parts:
- static data represent the authentication information (
login
, password) and the
user's personal data (surname, name, place of work, etc.). These data come from the
LDAP directory where MASC saves the user's information when he registers;
- the weighted data models the preferences of the user in relation to transport
criteria (least connections, the quickest, least amount of walking, the cheapest).
Each criterion is associated with a mark between 0 and 10, which represents the
degree of preference of the user in relation to this criterion;
- the history keeps track of the user interactions with PISs. This can act as a
knowledge base for the updating of static and weighted data. Analysis of the
information can, for example, inform the system that the user lives in Valenciennes
and works in Lille (as the user leaves for Lille in the morning and comes back to
Valenciennes in the evening, except on bank holidays). By analyzing the itineraries
chosen by the user, the system can deduce the preferences of the user in relation to
the modes of transport.
This user model is stored with the profile management agents in the form of an
XML document (an example is available in [ANL 06a]). Figure 3.20 presents the
activities carried out during the execution of the internal action
MAJPreferences
for
the skill
SaveChoiceSkill
. The objective here is to deduce the preferences of the user
according to criteria associated with the itineraries (least connections, the quickest,
the least amount of walking, the cheapest).
Figure 3.21 presents the activities carried out during the execution of the
getPreferedResponse
service of the
SocialFilteringSkill
skill. To select the itinerary
that is likely to interest the user, a majority vote (select the itinerary that was the
most chosen by users) is carried out on the itineraries if the current user has no
profile. If the user possesses a profile, and has already made his request, the
itinerary that he chose will be recommended. If the user has a profile but has never
made a request, a collaborative filtering
3
(it is the collaborative filtering method
based on the preferences and the behaviors of user which is applied) is carried out to
choose which itinerary to propose. This model therefore combines a cognitive
method (recommendation in relation to the profile) and social methods (majority
vote and collaborative filtering).
3 A synthesis of collaborative filtering techniques is available in [SU 09].
