The 8th International Conference on Computer Supported Cooperative Work in Design Proceedings

Collaborative Shopping Based on Multi-agent in Virtual Environments

2)ZhigengPan, ''Bing Xu, 'IMingming Zhang, 'IHongwei Yang ''State Key Lab o CAD&CG, Zhejiang University, Hangzhou, 310027, P.R. China f 2, Institute o Virtual Reality and Multimedia, HZIEE, Hangzhou, 31 0037, P.R. China f (zgpan,xubin,zmm,yanghongwei)@ad.zju. edu.en Abstract
Existing E-commerce applications on the Internet provide the users a relativeIy simple, browser-based interface to access available products. Customers are not provided with the realistic shopping experience as they would in an actual store or mall. So this paper presents a collaborative shopping system based on multi-agent that can provide the customers with simulation o the shopping experience in a real world. f With the creation and application o the avatar in the f 3 0 virtual shopping mall, the customers can change their role from "hollow-man" to human. By the f combination o the sociality with virtual environment, the system can not only imitates realistic communication among special customers who are favorable to the same products but also realize the communication between multiple avatars. Our collaborative shopping system creates the same virtual situation we are used to in real world: multiple customers can join and do the shopping together if they found that they are looking for the products with the same products area, they would chat with each other, ask the others' suggestions and find the desired products more effectively. The remainder of our paper is organized as follows: after reviewing the previous work related to our research, we give an overview of the concept in Section 3. In Section 4 describes the fundamental technology in our collaborative shopping procedure. Implementation of collaborative shopping in EasyMall is presented in Section 5. Finally, we summarize the contributions with fiture research in Section 6 .

2. Related Work
There has been a tremendous amount of previous work on creating a collaborative work and interaction spaces across networks, with notable contributions from the fields of Computer Supported Collaborative Work, Groupware, and Computer Human Interaction [Zj. In this section we briefly mention the previous work that come closed to our work. Xiaojun Shen [l] presents vCOM, a VRML and Java3D-based prototype, which permits users to navigate around virtual e-commerce worlds and perform collaborative shopping and intelligent searches with the assistance of software agents. Real-time interactions between the entities in this shared environment are implemented over the High Level Architecture (HLA), an IEEE standard for distributed simulations and modeling. Chia-Hui Wang [3] introduced a Web-based information system for providing remote surveillance services. The surveillance data can be treated as a kind of digitized product in E-market, which collaborative different value-added services with the same surveiIlance data. S. Puglia [Z] proposed a componentbased architecture for collaboration that provides shared navigation of the WWW along with an EJB-based server implementation. As a particular application built on this architecture, they present MuItECommerce, through which multiple users can participate in virtual shopping trips among multiple shopping sites. Promondia [4] provides a client-server architecture, based on the distributed of appfets, to support
386

1. Introduction
With the rapid expansion of the Internet, electronic commerce (E-commerce) using the web becomes popular. Nevertheless, existing E-commerce applications on the web provide the users a relatively simple, browser-based interface to access available products. The customers are mainly kept separated and everyone is shopping, as if hefshe was in an empty shop. Thus, customers are not provided with the same shopping experience as they would in an actual store or mall [I]. In our opinion collaborative shopping is something people like to do along with relatives and friends. In particular, it is likely that shopping is a social activity, meaning that the multiple customers can join the collaborative session to share with each other. We present the EasyMall system simulating most of the actual shopping experience through implementing the collaborative shopping based on creation of the multi-agent model in a virtual shopping mall environment. Multi-agent can search and recommend products according to the customer's preference. Using an avatar chosen from a vast range of avatar identities, the customers can walk around the virtual environment, look over. and manipulate the products that he is interested in and order goods through a security system.
0-7803-794 1-1/03/$17.00O2003 IEEE.

collaborative tasks such as text-based chat, shared whiteboards, voting and surveys and implements a system where the interaction is limited to one Protnondia server distributing page and services. Our EasyMall system is a multi-server based on agents [ 5 ] , which uses blaxxun to implement the interaction between customers and the virtual environment. Its goal is to create an interactive virtual mall with integrated Ecommerce, agent technology and virtual reality. VRML and Java3d technology is employed to provide avatar for each customer who can communicate with other avatars as well as products in mall and agents [ 6 ] .

3. System Overview
We brought forward the EasyMall system based on the client-server architecture to develop and implement a collaborative shopping session. Basically, the collaborative shopping system can be divided into four layers as shown in Figure 1.

consecutive high-level processes: multi-agent and collaborative management. In the multi-agent level, there are three agents: Recommendation agent, database agent and manipulation agent. A detailed description will be showed in Section 4. The collaborative management level is also included in the layer, which is composed of object ownership administrator, security and scene control. In Section 4, we will discuss these in detai 1. Server Layer: the server layer includes multiple merchant servers. Every merchant server stores the various products and merchants information. The agent engine layer communicates with the server layer in xml and accesses the relative data.

4. Collaborative Shopping based on Multiagent

In this section, we introduce the core module of the system. We start with explaining the main components of the multi-agent. Then, we describe the collaborative management and its components.

4.1. Multi-agent Mechanism

To offer accurate and required products which the customer wants to purchase, the EasyMall system used multi-agent which can provide the customer with selection and manipulation of the products according to customers' preference. During the procedure, database agent is also used to manage and manipulate varieties of data such as products database, sales database and customer. profile data. The framework of multi-agent is showed in Figure 2 .
I
I I I I

Java3DNRML-based browser

I ,

Figure 1. An overview of the EasyMall system architecture

Shared+

"ern

I I
I

Client Layer: as for the distribution of the shopping, eveiy user in the client layer may collect items from the different merchant sites he is interested in. The location the user is visiting, expressed in terms of the UlUs of the 3D virtual environment he currently navigates from his terminal. Interaction Layer: at client site, the interaction layer provides the interaction interface for every user. The browser based on Java3D or VRML mode offers some performance advantages to users, such as choosing the avatar, chatting between avatars and controlling the roam of avatar, etc. Agent Engine Layer: the agent engine layer plays a pivotal role, which supports the communications of the collaborative shopping in virtual environments. The agent engine layer can be seen as including two
387

I
I

!
Figure 2. Multi-agent structure

In the following section we describe how the multiagent guides customers to purchase.

(1) Database agent In current E-commerce, the most popular data model for a data warehouse is a multidimensional model. Such a model can exist in the form of a star schema, a snowflake schema, or a fact constellation schema.

Figure 3 shows the star schema which is used in EasyMall system. Purchasing is considered along four dimensions, namely, products, time, customer, and merchant. The schema contains a central fact table for purchase that contains keys to each of the four dimensions. To minimize the size of the fact table, dimension identifiers (such as time-key and merchant-id) are system-generated identifiers [81.
Dimension tables Customer table
uataner k g r r m e custaner ap ulStaner gender custaner Mty

Fad table Purchase table

Dimenson tables m e d a o t table

merchant type merchant raig

I -

Although a domain expert can design the text dialogs, the best methods to show his individual preferences can be expressed by the customer. Decision tree induction can be used for attribute subset selection. A tree is constructed from the given data. All attributes that do not appear in the tree are assumed to be irrelevant and can be deleted. Rough set can also implement this goal. Here, we present an approach to feature extraction in which feature selection, feature extraction, and classifier training are performed simultaneously by the integration genetic algorithm and case base reasoning. Figure 4 shows the flow chart of the algorithm.

produd table

time table

Feature masurem

Feature vector

Feature extraction

Figure 3. Star schema of a data warehouse for purchase Figure 4. The flow chart of the algorithm

Notice that in the star schema, each dimension is represented by only one table, and each table contains a set of attributes. For example, the customer dimension table contains the attribute set (id, name, age, gender, hobby}. In order to divide customers into different groups and recommend products to target group, we design Database agent based on the star schema so that it can satisfj, the need to understand the behavior of business units such as customers and products. Database agent stores all information concerning the user profile, the purchase history records, virtual environment including characteristic of avatars, the history of their encounters with other avatars and the click of user, so that the users can not only be recommended the favorable products but also be presented with information about the current situation of and encounters made by the avatar in the virtual mall when the user logs in and chooses hisher avatar.

The first step: Through the questionnaire, we obtain the feature vector. 0 The second step: using GA to select the features. 0 The third step: generating new feature vector and each feature has different weight. 0 The fourth step: we use CBR as the classifier. The CBR classification has been selected for use in combination with the GA feature extractor for two reasons. The simplicity of the CBR classifier makes it easy to implement. Weights used in computation of similarity can represent each features importance, which can improve prediction accuracy. In the following, we describe the combination of GA and CBR technology in detail.
0

Genetic algorithm

(2) Product selection agent

When the customer enters the EasyMall, he first inputs his wishes or demands. But if the system asks the customer to input a lot of require information in which customers are not interested, it will not only guide the system to find many irrelevant products but also annoy the customers and finally lead them to log out the system. To avoid this situation, Product selection requires that products be described in terms of a set of measurable features that the customers are concerned about. The selection and quality of the features representing each customers requirement have a considerable bearing on the success of subsequent selection.
388

Algorithm GA: Genetic Algorithm Step 1: Initialize a population of chromosomes Step 2: evaluate each chromosome in the population Step 3: create new chromosomes by mating current chromosomes; apply mutation and recombination as the parent chromosomes mate. Step 4: delete members of the population to make room for new chromosomes. Step 5 : evaluate the new chromosomes and insert them into the population. Step 6: if the stopping criterion is satisfied, then stop and return the best chromosome; otherwise, go to step 3.
Chronxmrne i
000000000

I 101111010 I 000000000 I ... I 110001101

N*(l+8) bits

N-dimensional vector comprises a single member of the GA population for GA-based feature selection. Each vector has 9 bits. The GA feature extraction technique has been expanded to include a binary masking vector along with the feature weight vector on the chromosome. If the mask value for a given feature is zero, the feature is not considered for classification such as feature 1. If the mask value is one, the feature is included in the classifier such as feature 2 and its weight w 2 is 01111010 Since the particular GA engine used maximizes the objective function, the following object function is used to evaluate each chromosome: Object function = No. of masked features + No. of correct predictions During the process, presuming that only those chromosomes that can classify the training set with higher accuracy than the given threshold can join into the next population and those chromosomes below the given training accuracy can be deleted.
Case Base Reasoning

skirt, the product manipulation agent can help her to change the texture based on the texture morphing method [9] (depicted on Figure 5).

Figure 5. The simulating of skirt

4.2. Collaborative Management

The collaborative management is the central level of the collaborative shopping system. The current virtual market places often lack in the emulation of the social interaction factors [7]. So in our system we combine the sociality with virtual environment. On the first log in, the user chooses hisher avatar representative by registering information; therefore, the customers change their role from hollow-man to human. The customers might walk and select products in the virtual environment by avatar, they can also invite other avatars to meetings with ones own avatar or with accepting invitations for the own avatar. The collaborative management allows the customers to manipulate the products in natural manner such as one-customer selects T-shirt from the costume area. In the following section, we will introduce the components of collaborative management.
4.2.1. Object Ownership Administration

The intelligent assistant uses the GA technology to capture the customers demands and wishes. Then based on this information, the intelligent assistant converts the inaccurate information into fuzzy variables in terms of its membership functions and then uses Case Base Reasoning (CBR) methods to choose the appropriate products. CBR classifiers are instanced-based. When given a new case to classify, a case-based reasoning will search for training cases having components that are similar to those of the new case. The key technology in CBR is finding similarity measure. Here, a query q is described by the attributes q ,q ,... q II and a case c is c described by c 1 , 2 ,... C , , the similarity defined as
(4 > c ) =
0

can be

2
,=I

WicT i

Where the weights

w, and q , can be acquired by

GA.
(3) ]Product manipulation agent

As to the manipulation agent, it allows customer to conlrol an avatar to do some motions so as to inspect the products he customized through the animation playing with a third person view point or operate products with the first person viewpoint. Product manipulation agent is also used to control the behavior of objects. When the avatar picks up the products, he can look over the products from different angles and can also read the relevant information about the products through text, image and video. For example, when an avatar enters the clothing location, she can select the favorable product and try it on. If she is not satisfied with the color or texture of the
389

The ownership of object is one of the pivotal models in the procedure of collaborative shopping. The collaborative management module provides the ownership core based on mutex lock mechanism for exchanging attribute ownership among multiple avatars in the shared virtual space so as to ensure that only one avatar at a time has access to the product. Pthread-a mutex; //definition o the mutex f semaphore Pthread-mutex-init(mutex, mutexattr); // initialize the mutex variable, the first parameter is the variable that will be initialized, the second parameter i the s attribute o the mutex variable, f Pthread-mutex-lock(mutex); // lock the mutex variable Pthread-mutex-unlock(mutex); // unlock the mutex variable

The ownership core uses the above sub-function to administrate the object ownership, which permits only one avatar to handle the object, so multiple avatars exclusively manipulates the attributes of an owned object (such as orientation and location). If another avatar attempts to operate the object whose ownership has been distributed, the avatar is inserted in the waiting queue until the ownership of object was released. For instance, when an avatar clicks the T-shirt that he wants to operate, he/she first needs to request the ownership of object by using the ownership core. If the ownership of the T-shirt has already been distributed, the avatar has to wait until the ownership is given back. Then, ownership core allots the ownership to other avatars who are in the waiting queue according to the priority and waiting time. After that he/she can put the T-shirt back by releasing the ownership to its original owner.
4.2.2. Security

The security, another pivotal model, is responsible for controlling the privacy of customers and merchants. In fact, the collaborative shopping in virtual environment is done through the interaction of remote object between the client layer and server layer. In the system, we use Java and Java3D to implement the object remote control. The problem we encountered is due to restrictions imposed by the Java security model on remote access to an applet. Firstly, the absence of support for multiple inheritances in Java made it impossible to have an applet, which already inherits from the Java Applet class, also extend the Unicast Remote Object (which is necessary to make an object remote). Secondly, there seem to be server security obstacles preventing the server layer from opening a separate socket to actively write on a remote object bound to the applet. The problem is that an applet may not open a server socket to listen for requests from the arbitrary hosts [ 2 ] . As to privacy, we used the security model to provide the mechanism to degrade or enhance the customers and merchants privacy. On the one hand, the security model enables the possibility to provide anonymity to customers since it could prevent cookies and other private data from being transmitted to the merchants, acting as a filter between the two. On the other hand, the information that is available to the security model relating to customer identity, preferences and shopping habits would be of great value to merchants and vendors.
4.2.3. Scene Control

the system [lo]. Especially, in large virtual worlds the number of objects that require certain synchronization or update messages to be transferred over a network can slow down the interaction of the individual user with the shared world in an unreasonable way [ 2 ] .A solution for this problem is the subdivision of large virtual worlds into several regions or zones [l 11. For the collaborative shopping environment, we partition the virtual environment into several separate areas. Each of the areas in the virtual environment has its own boundary and world coordination. In our approach, we use the Proximity Sensor of VRML to listen the avatars coordination. The coordination is represented by [e,X , Y ,21, where Q represents the ID of certain shopping area which the avatar is in, ( X ,Y ,Z)represents the avatars position information in the shopping area. For example, when an avatar walking inside the boundary of T-shirt area, the sensor transmit the information to scene control model, so it refreshes and synchronously updates the corresponding data on the participating avatar.

5. Implementation Shopping

of

Collaborative

We implemented a multi-user collaborative shopping virtual environment by using VRML, Java3d, artificial intelligent and computer network technology. In our system, we provide the user with simple and efficient navigation tools to visit the multiple virtual malls (depicted in Figure 6). The EasyMall system provides the multi-agent model to allow the avatar communicate with the intelligent guider. When customers enter the virtual mall, the intelligent guider help customers find out what they really want. When the customers can simply identify the type of product they need and describe the features or specific functions of the products.

The traditional distributed system is based on a central server model. In this model clients communicate with a central server, which manages the entire system and informs clients of any updates and changes in the state of avatars and objects. Clients only communicate with the stand-alone server, which contains the entire scene database and tracks all objects of interest within
390

Figure 6. Multiple virtual malls (Entrance, Products Selection Area, Garment Area and Furniture Area)

Customers might interact with other avatars in the shopping procedure. If two avatars meet in a T-shirt

more effectively help the customers find and choose the desired goods [2]. Figure 7 shows an example in the EasyMA system where two avatars collaboratively buy T-shirt. The female avatar finds a beautiful T-shirt, shows it to the male avatar and asks his opinion.

which can improve the performance of collaborative shopping.

7. Acknowledgement
This project is co-supported by Excellent Young Teacher Award Project of MoE, 973 Project (grant no: 2002CB312100) and European ELVIS Project.

8. References
[11Xiaojun Shen, T.Radakrishnan, Nicolas D.Georganas.

VCOM:Electronic commerce in a collaborative virtual world, Electronic Commerce Research and Applications l(2002) 281-300. [2] S . Puglia, R. Carter, and R. Jain, MultECommerce: A
distributed architecture for collaborative shopping on the www, in ACM Conference on Electronic commerce,

2000.19. [3] Chia-Hui Wang, Ray-I Chang, Jan-Ming Ho, An Effective Communication Model for Collaborative Commerce of Web-Based Surveillance Services, IEEE International Conference on E-Commerce, 2003. [4] Gall U., Hauck F.J., Promondia: A Java-Based Framework for Real-time Group Communication in the Web,
Proceedings of the Sixth International World Wide Web Conference (1997). [5] Zhigeng Pan, Bing Xu, Tian Chen, Fangshen Wu. Design and Implementation of an Interactive 3D Virtual Shopping System EasyMall. Journal of CAD&CG (To appear). [6] Qingge Ji, Bingrong Hong, Dongmu Wang. Mobile
Figure 7. Collaborativeshopping in a virtual T-shirt store

6. Conclusion and Future Work

We have presented Easymall, a system aimed to support collaborative shopping based on multi-agent in virtual environment. With the creation and application of the 3D virtual shopping mall, the customers might simulate the shopping experience in a real world by avatar. We have integrated virtual reality and intelligent technology into E-commerce to generate a collaborative shopping prototype. The prototype not only imitates reapistic communication among special customers who are favorable to the same products, but also provides human-like interactive interface. Our future work will design a virtual memory with the smart object technology to implement multiple avatars autonomic behavior in virtual environment,

Agent based Prototype of Heterogeneous Distributed Virtual Environment Systems, Journal of Systems Engineering and Electronics, 2000, ll(2): 61-65. [7] T. Schuemmer, GAMA-Mall-Shopping in Communities, in: Second International Workshop on Electronic Commerce (WELCOMOl),Heidelberg, 2001. [8] Jiawei Han, Micheline Kambr Data Mining Concepts and Techniques Morgan Kaufmann Publishers. [9] P.Wang, M.M.Zhang, Z.G.Pan, New texture morphing method for visual presentation of textile product, Second International Conf. on Image and Graphics, Aug.2002, HeFei,China,PP1011-1016. [IO] P. Morillo, M. Fernandez, N. Pelechano: A Grid Representation for Distributed Virtual Environments. 2003 Annual Crossgrid Project Workshop & 1st European Across Grids Conference, February, 13th-I4th, 2003. [ll] W.Brol1, Populating the Internet: Supporting Multiple Users and Shared Applications w t VRML, in: ih Proceeding of the VRML97 Symposiun, (Feb. 24-28, 1997, Monterey, CA), ACM SIGGFLG, ACM, 1997.pp.87-94.

391