A Genetic Fuzzy Decision Agent based on Personal
4.
Ontology for Meeting Scheduling Support SystemT
Chang-Shing Lee Meng-Ju Chang
Management Management
Department of Information Chang Jung Universiiy
Tainan, 711, TAIWAN
leecs@mail.cju.edu.tw 176901 106@maill .cju.edu.tw Abstract
-
A Genetic Fuuy Decision Agent (GFDA) based on personal ontology for Meeting Scheduling Support System (MS3) is proposed in this paper. We apply the concept of personal ontology to the MS3, and extend the four-layered object-oriented ontology to Personal Meeting Scheduling Ontology (PMSO) for the MS3. When an organization member sends a meeting request with required information to MS3, the MS3 will retrieve the invitees’ information from Personal Mepring Scheduling Ontology(PMSO) Repository and send a response to GFDA. Then, GFDA utilizes these invitees’ information along with the PMSO and meeting information to infer the suitable meeting time slots. Therefore, GFDA can analyze each invitee’s attendance possibility based on soft-computing technology. The experimental result shows that the proposed GFDA is feasible, efficient and usable for meeting scheduling support system.
I. INTRODUCTION
Ontology is a collection of key concepts and their interrelationships collectively providing an abstract view of an application domain. With the support of ontology, both user and system can communicate with each other by the shared and common understanding of a domain [I]. Moreover, an ontology is a computational model of some portions of the world. It is often captured in a semantic network and represented by a graph whose nodes are concepts or individual objects and whose arcs represent relationships or associations among the concepts [2]. On the other hand, an agent is a program that perform unique tasks without direct human supervision. I. Ferber [3] gives another definition of an agent such as “an agent is capable of acting in an environment and can communicate directly with other agents”. An intelligent agent is more powerful than an agent because of the reasoning and learning capabilities [4]. Employees in an enterprise may spend much of their time scheduling and attending meetings. The process of searching for an available meeting time can be complicated by communication delayed and by other concurrently scheduled meeting. Automating meeting scheduling not only can save the user time and effort, but also lead to more efficient schedulers and improvements in how the enterprise exchanges information.
S . Sen [SI proposes a system using intelligent meeting- scheduling agents that can negotiate with other agents without
Department of Information Chang Jung Universiiy
Tainan, 711, TAIWAN
Chen-Yu
Pm
Chyi-Nan ChenDepartment of Engineering Department of
Science Engineering Science
National Cheng Kung Tainan, 701, TAIWAN n9692412@ccmail.ncku.edu.tw cnc@mail.ncku.edu.tw
compromising their user-specified constrains. In addition, K. Sugihara et al. [6] propose a meeting scheduler for ofice automation. They took the priorities of persons and meetings as considerations in the real office environment, and use a heuristic algorithm for timetable rearrangement. T. Haynes et al. [7] propose an automated meeting scheduling system that utilizes user preferences. W. S. Jeong et al. [8] focuse on bow the meeting-scheduling agent can reduce failures when there is no common time-slot. They solve the failure condition with utilizing the cooperation and the rescheduling strategy. E. Mynatt et al. [9] present a calendar system extension that uses a Bayesian model to predict the likelihood of one’s attendance at the events listed on his or her schedule. A. Ashir et al. [IO] propose a multi-agent based decision mechanism for distributed meeting-scheduling system. J. A. Pino et al. [I I] accommodate users’ availability according to their own preferences and restrictions of schedule meetings. F. Bergenti et al. [I21 describe an agent-based computer-supported cooperative work system designed to promote the productivity of distributed meetings by means of agent. C. Glezer [I31 proposes and evaluates a comprehensive agent-based architecture for an inter-organizational intelligent meeting- scheduler. C-S Lee et al. [I41 use an intelligent fuzzy meeting agent to infer attendance possibility. In this paper, we develop a Genetic Fuzzy Decision Agent (GFDA) based on personal
ontology for Meeting Scheduling Support System (MS3).
This paper is organized as follows. In Section 11, we apply the concept of personal ontology to meeting scheduling support system, and use a four-layer object-oriented ontology for this application. Section I11 introduces the architecture of meeting scheduling support system. The experimental results are presented in Section IV. Finally, the conclusion is given in Section V.
National Cheng Kung Toinan, 701, TAIWAN
Universiiy Universiw
a.
PERSONAL ONTOLOGY FOR MEETING SCHEDULINGSUPPORT SYSTEM
With an ontology, we can organize keywords and database concepts by capturing the semantic relationship among the keywords or among the tables and fields in databases [2]. The semantic relationship can provide an abstract view of the information space for our schedules. T. R.
This work is partially supported by National Science Council of Republic of China under grants NSC92-2213-E-309-005
105
R a p e et al. [15] propose a distributed meeting scheduling agent, called RETSINA Calendar Agent (RCal), that processes schedules marked up on the Semantic Web, and imports them into the user’s personal information manager. In this paper, we apply the concept of personal ontology [Z] to the meeting scheduling support system, and extend the four- layered object-oriented ontology [ 161 to Personal Meeting Scheduling Ontology (PMSO) for the MS3.
In the proposed ontology, we d e f i e four layers including Domain Layer, Category Layer, Class Layer, and Instance Layer, and four kinds of relationships including association, generalization, aggregation, and instance-of for meeting scheduling application domain. The association relation belongs to non-taxonomic relation. On the other hand, the generation, aggregation, and instance-of belong to taxonomic relation. Besides, the aggregation is a whole-part relationship.
\
-
I.--Fig. 1 The architechxe of Penanal Meeting Scheduling Ontology. Fig. 1 shows the architecture of Personal Meeting Scheduling Ontology (PMSO). In the MS3, the Category Layer is classified into “Teacher Category”, “Research Assistance Category”, and “Student Category”. Fig. 2 shows a part of Personal Meeting Scheduling Ontology (PMSO) for “Decision Support & Artificial Intelligence (DSAI)” Lab. at Cbang Jung University in Taiwan. There exist semantic relations between the concept pairs in the Instance Layer. For example, there are two Concepts “Cbang-Shing Lee: Teacher” and “Meng-Ju Chang: Postgraduate” in Instance Layer. The association relationship, “Participate” and “Guide”, are existed between the concept pairs, (Teacher, Department Meeting) and (Teacher, Postgraduate), respectively.
Fig. 2 A part of Penanal Meeting Scheduling Ontology for DSAI Lab.
HI. THE ARCHITECTURE OF MEETING SCHEDULING SUPPORT
SYSTEM
A. The Architecfure of Meeting Scheduling Support System
GFDA j
~ ...
Fig. 3 The arehitechrre ofmeeting scheduling support system.
Fig. 3 shows the architecture of MS3. At the beginniig, the meeting host sends a meeting requirement to MS3, and MS3 can obtain the invitees’ schedules from the Personal Meeting Scheduling Ontology (PMSO) Repository. MS3 then computes the common free time or removable working time for all invitees, and responds the computing results to Parallel Fuzzy Inference Engine (PFIE). The PFIE performs fuzzy inference by utilizing the invitees’ information, the fuzzy knowledge base of PMSO, and meeting information. In this architecture, the devices of end users include desk computer, cell phone, notebook, PDA, etc. Hence, the invitees can get the meeting information by various platforms. Furthermore,
the attendance records of the invitees will be keep in the PMSO for generating invitee's personal fuzzy knowledge base. In the specific time, Genetic Learning Agent (GLA) will start to learn the fuzzy knowledge base for each user. In addition, an evaluation module is used to evaluate the learning process. The Genetic Fuzzy Decision Agent will be introduced in next section.
B. Genetic Fuzzy Decision Agent
The Genetic Fuzzy Agent consists of a parallel Fuzzy Inference Engine, a Genetic Learning Agent, an Evaluation Module and a Personal Meeting Scheduling Ontology Repository. We utilize a Parallel Fuzzy Inference Engine to infer attendance possibility of each invitee (See Fig. 4)
* m b c s w * t L ~ [141[171.
Fig. 4 The architechre of Parallel F u u y Inference Engine.
I ) Parallel Fuzzy Inference Engine:
There are seven input fuzzy variables used in the Parallel Fuzzy Inference Engine for MS3. The User-Priori@ (UP) fuzzy variable denotes the importance of invitees. The Meeting-Event-Priori@ (MEP) fuzzy variable denotes the importance of the meeting. The length of meeting time is
denoted as the Meeting-Time-Length (MTL).
Meeting-Place-Preference (MPP) fuzzy variable denotes the preference of meeting place for each invitee. The M e e f i n g - S u b j e c f n c e (MSP) fuzzy variable denotes the preference of meeting subject for each invitee. And the last two fuzzy variables are Meeting-Time-Preference-1 (MTPI), and Meeting-Time-Preference-.? (MTP2). MTPl ficzzy variable denotes the preference of meeting time for each invitee. For example, if one people likes to attend the meeting in the moming at AMlO:OO, and the meeting time is AM9: 30, then PFIE may infer the possibility of attending this meeting for the invitee is high. MTP2 fuzzy variable is used to consider the work priority of each invitee's schedule. For example, if a meeting event priority for user A is low, and A has a high priority work to do at the same time, then A may not attend the meeting at this time. The output fuzzy variable in the PFIE is Attend-Meeting-Possibility (AMP), it denotes the possibility of attending the meeting for each invitee. Fig. 5
shows an example of membership functions for fuzzy variable UP.
Membership
User Priorip
Fig. 5 An example of membership functions for fuzzy variable UP. 2) Genetic Learning Agent:
We modify the genetic learning method [18] and data- driven method [19] to generate the fuzzy knowledge base. The knowledge base includes the data base @B) and rule base
(REI).
The DB consists of the number of linguistic terms and the parameters of membership functions of each fuzzy variable UP, MEP, MTL, MPP, MSP, MTPI, MTPZ and AMP. The RE3 consists of fuzzy rules.In the beginning, we encode the DB to be composed of two parts in each chromosome. One is the number of linguistic terms c,, and the other is the parameters of membership functions c, . The number of linguistic terms for each variable is stored into a vector c, as follows:
c, = @,, L,. ... , LJ (1)
where L, represents the number of liguistic terms of i-th
fuzzy variable. The value of each L, is restricted in the set (2, 3, 4 ) . On the other hand c, represents the parameters of each membership function. Fig. 6 shows the membership h c t i o n s used in Genetic Learning Agent.
xi
O Pd p,,
...
p,,'
Fig. 6 The membership functions used in Genetic Learning Agent.
The parameters of membership function for each variable
c, =(C,,. c,. ...
.
C>J (2) where c,, represents the parameters of membership(3) where p , represents the peak value of j-th linguistic term in the i-tb fuzzy variable. Therefore, the chromosome are stored into a vector c, as follows:
functions of i-th fi(.zy variable, and
c,,
=e,,
ea,
... , P,) 3 andel
<e2
< ... < P,c = c,c, represents to the membership functions of each fuzzy
variable.
In addition, we utilize data-driven method to generate
fuzzy rules from attendance records in PMSO and
chromosomes. The attendance record is the set of input-output data pairs as follows:
(4)
where denotes the input values of i-th fuzzy variable for the j-th attendance record. denotes the actual decision for thej-th attendance record. Each data pair will generate a fuzzy rule, but it is highly possible that there will generate some conflicting rules. Therefore, we assign a degree to each rule and take the j-th rule with maximum degree D(Ru1e‘) to resolve conflicting rules as follows:
D(Rule’) = p ( x : ) x p ( x ; ) x ... x p ( x ~ . , ) x p ( Y i ) (5)
3) Evaluarion Module: the fitness function as follows:
In this paper, we use the mean square error (MSE) to be
where T denotes the number of the training data, y ’
denotes the output of PFIE for the :-th training data, and yk denotes the desired output of the t-th training data.
Next, we briefly describe the related genetic operators used in this paper as follows: (a) Selection: The selection probability calculation follows linear ranking. The selection probability is computed by using the nonincreasing assignment function; (b) Crossover: The standard crossover operator is applied over the two parts of tbe chromosomes. When c, is crossed at a random point, the corresponding values in c, are also crossed in the two parents; (c) Mutation: Two different operators are used in c, and c,
,
respectively. Inc,
part, we random select the number of linguistic terms of the variable and change it to the immediately upper or lower value. In c, part, Micbalewicz’s nonuniform mutation operator [20] is used.IV. EXPERIMENTAL RESULTS
In order to evaluate the effectiveness of the proposed GFDA on MS3, we adopt the actual attendance records retrieved &om PMSO repository as the experimental data. We evaluate the efficiency of genetic learning in the fust experiment and adopt the latest 100 attendance records of five Lab. members. Then, 80 and 20 attendance records are adopted as the training and testing data respectively. Table I
demonstrates the computing results. The defmition of each column for Table I is described as follows:
I ) User: the member of DSAI Lab.
2)Granularity: the number of labels for each fuzzy
3)Rules: the number of rules.
4) MSE,,,: MSE over the training data set. 5) MSE,: MSE over the testing data set.
variable.
TABLE I
T H E C O M P m G RESULTS OF 5 MEMBulS IN DSAI LAB
4 2 3 3 3 4 3 4 0.004999 0,004000 4 4 2 3 4 4 2 4 0.008749 0.037133 3 4 2 2 2 4 2 2 0.005536 0.029250 5 3 3 3 4 3 3 3 4 65 0.008808 0.041291
In the second experiment, we adopt one of the attendance records fiom DSAI Lab. members. Whenever the MS3 increases five attendance records of the member, GLA will start to learn the Furzy Knowledge Base of personal ontology. Then, the Fuzzy Knowledge Base generated by GLA will assist PFIE to process more correct inference for the next meeting.
As shown in Fig. 7, the experimental result illustrates the frequency of the correct inference after learning. Obviously, the more attendance records that increased in the MS3, the more correct inference can be inferred by the GFDA. In this experiment, the times of correct inference are 83, and the total times for fuzzy inference are 95, hence the average correct rate is 83/95=87.37%. This experimental result shows that the GFDA can work effectively.
t 2 3 4 5 6 7 8 9 10 I I I2 I > I4 IS 16 I7 18 I9
l“g Tvor.
Fig 7 The H i s t o g m of Times of Correct Inference.
V. CONCLUSIONS
A Genetic Fuzzy Decision Agent based on personal ontology for Meeting Scheduling Support System is proposed in this paper. The experimental results shows that the GFDA is feasible, efficient and usable for meeting scheduling support system. In the future, we will extend the capability of GFDA to be able to generate the semantic meeting information by personal ontology.
REFERENCES
[I] V-W So0 and C-Y Lin, “Ontology-based information retrieval in a multi-
agent system for digital library,” Proceedings ofthe 6* Conference on
[2] M. N. Huhos and L. M. Stephens, "Personal Ontologies," IEEE Internet
Computing,vol. 3, no. 5,pp. 85-87, Sep./Oet. 1999.
[3] I. Fcrber, Mulli-Agent Systems, ADDISON- WESLEY, New York, 1999.
[41 H. Chen, A. Hausian, I. Nunamaker and J. Yen, "Toward Intelligent
Meeting Agents," IEEE Computer, vol. 29, no. 8, pp. 62-70, Aug. 1996.
[SI S. Sen, "Developing an Automated Distributed Meeting Scheduler,"
IEEE ExperLvol. 12, no. 4, pp. 4 1 4 5 , JulylAug. 1997.
[6] K. Sugihara, T. Kikuno and N. Yashida, " A Meeting Scheduler for office
Aulomation," IEEE Tram. on Sofioore Engineering, vol. 15, no. IO, pp.
1141-1146,Oct. 1989.
171 T. Haynes and S. Sen, N. &ora and R. Nadellq "An automated meeting
scheduling system that utilizes user preferences," Proceedings ofthe I"
inlemationol conference on Autonomous agents, USA, 1997, pp. 308-
315.
W. S. Jeong, J. S. Yun and G. S . Ja, "Coopcration in Multi-Agent
System for Meeting Scheduling," Proceedings of the Region 10
conference on IEEE TENCON, Korea, 1999, vol. 2, pp. 832-835.
[9] E. MynaU and J. Tullio, "Inferring Calendar Event Attendance,"
Proceedings
sf
the rSh lntemational Conference on lnlclligent urerlnte$oce~,USA, 2001,pp. 21-128.
[IO] A. Ashir, K. H. loo, T. Kinoshita and N. Shiratori, "Multi-Agent Based
Decision Mechanism for Distributed Meeting Scheduling System,"
Proceedings of the lntemationol Conference on Porollel and Distributed Syrlems, Korea, 1997, pp. 275-280.
[ I l l J. A. Pino and H. A. Mora, "Scheduling Meetings with Guests'
approval," Proceedings of the Wll Intemationol Conference of the
Chileon on IEEE Computer Science Sociev, Chile, 1997, pp. 182-189. [I21 F. Bergenti and A. Paggi, "An Agent-Based Approach to Manage
Negotiation Protocols in Flexible CSCW Systems," Proceedings o/the
lh lntemotionol Conference on Autonomous Agenh, Spain, 2000, pp.
267-268. 181
[I31 C. Glezer, "A conceptual model of an interorganizational intelligent
meeting-scheduler (IIMS)," Journal of SIloregic Infomotion System,
vol. 12, no. I, pp. 47-70, March 2003.
[I41 C-S Lee and C-Y Pan, 'An Intelligent F u u y Agent far Meeting
Scheduling Decision Support System" Fuay Sets ondQstems, vol. 142,
no. 3, pp. 467488,2004.
[IS] T. R. Payne, R. Sin@ and K. Sycara, "processing Schedules Using
Distributed Ontologies on the Semantic Web," Workrhop on Web
Services, E-business and the Semmtic Web, Springer-Verlag Heidelberg, pp. 203-212.2OfJ2.
[I61 C-S Lee, Y-F Kao, Y-H Kuo and I-H Meng, "An Episode-based Fuzzy
Inference Mechanism for Chinese News Ontolow Conslruction," The f k
World Multiconzrence on Systemics, Cybe%icF and lnfimotics,
USA, 2003, pp. 453458.
1171 C. T. Lin, C. S. G. Lee, "Neural-Nelwork-Based FW Logic Control . .
and Decision System,*' IEEE Tram. on Computers, vol. 40,no. 12, pp.
1320-1336,Dec. 1991.
[I81 0. Cordon, F. Herrera, and P. Villar, "Gmemting the Knowledge Base of a F u u y Rule-Based System by the Genetic Learning of the Data
Basc," IEEE Tram. on Fwry Systems, vol. 9, no. 4, pp. 667-674, Aug.
2001.
[I91 L. X. Wang and J. M. Mendel, "Generating Fuuy Rules by Leaming from Examples," IEEE Tmm. Syrtem~ Mon. and Cybernetics, vol. 22, no. 6,pp. 1414-1427,Nov.iDec. 1992.
[20] 2. Michalewicz, Genetic Algorithms