-In this paper, we study the issue of peer to peer adaptive awareness. More specifically, we study how agents can facilitate and mediate interaction, communication and cooperation among people. We propose the concepts of a smart distance and an awareness network in a distributed collaborative environment. We illustrate the architecture of an Agent Mediated Collaborative system - the Agent-Buddy system that can create a sense of group presence and, at the same time, preserve the privacy of each user. Virtual springs systems are used to model the awareness degrees among team members. Each agent makes decisions by considering multiple factors. The goal of the multiagent team is to minimize the global awareness frustrations with respect to different kinds of tasks. Empirical studies have been conducted to analyze the influence of individual behavior on global performance for various kinds of tasks.
Tcompanies and institutions to de-centralize their
HE worldwide nature of today's market has forced many organizational structures. Furthermore, more and more people will be working from home. With ubiquitous connectivity on the horizon, collaborative computing promises to become one of this century's core applications. People will be more and more involved in collaborative computing because of the pressure from companies to improve their productdevelopment and decision-making processes and because of the convenience brought by the information super-highway.
There are four modes conceptualized by researchers in
Computer Supported Collaborative Work (CSCW) on how
people work in a collaborative environment [
Many computer systems support simultaneous interaction by
more than one user. However, most of them support multi-user
interaction in a way that prohibits cooperation – they give each
user the illusion that he or she is the only one using the system.
To support and encourage cooperation, cooperative
applications must allow users to be aware of the activities of
others. The purpose of providing cooperative awareness is to
establish and maintain a common context and to allow the
activities or events associated with one user to be reflected on
the other users' screens. For example, Lotus Sametime [
When more and more people are working in a distributed cooperative environment, especially when more and more people are working from home, the requirement of staying aware of co-workers’ status and activities will become increasingly important. Parallel with the advances made in CSCW in recent years, there have been interesting developments in the fields of Intelligent Agents and Distributed Artificial Intelligence, notably in the concepts, theories and deployment of intelligent agents as a means of distributing computer-based problem solving expertise. The concept of intelligent agents has given rise to an exciting new technology of wide-potential applicability. In particular, the paradigm of multi-agent systems forms a good basis for the design of CSCW architectures, and the support of CSCW operations. Intelligent agents that can undertake sophisticated processes on behalf of the user and dynamically and intelligently adjust the “distances” among co-workers will be a necessary part of any organization’s virtual structure. The digital multi-agent organization will capture the dynamics of teamwork, adjust the awareness level among co-workers, and re-shape the form and characteristics of collaborative work. The automation brought by Agent Supported Cooperative Work (ASCW) system will dramatically reduce certain types of frictional costs. On a larger scale, it is our belief that in the future, the WWW will not only be the knowledge pool of human society, but also be the digital world where people can meet and sense each other.
The remainder of this paper is organized as follows. The next section gives a definition of peer to peer computing. Section III proposes the concept of smart distance. Section IV describes the architecture of Agent-Buddy - an ASCW system that provides an adaptive awareness among co-workers. Section V defines the concept of an awareness network, which is a key concept behind Agent-Buddy. Section VI details the mechanism of adaptively adjusting the awareness levels in Agent-Buddy. Section VII empirically studies the influences of agents’ behaviors on global performances with respect to different kinds of tasks. Section VIII discusses related work. Section IX presents brief conclusions.
Peer-to-peer computing [
People are separated by distance and they like to adjust it when there are choices. For example, when working at the same table, the two persons in Figure 1(a) are quite close, while the two persons in Figure 1(b) are not so close. In Figure 1(c), physical rooms are built to separate co-workers. Technologies, however, can bring distant people closer, as shown in Figure 1(d).
Distance, as an abstract concept here, refers to the degree of objective difficulties in sensing other people through taste, touch, smell, hearing, and sight. Physical distance, as determined by the geometrical distance of body centers, is one of the major factors that determine the distance between people. However, it is not the only factor. Environment also contributes to the sense of distance. For example, occlusions can increase the difficulties in sensing, thus increasing the distance.
Technologies can provide more communication channels and thus shorten the distance. In a two-person telephone conversation scenario or video conferencing scenario, the distances between people are made much shorter because they can hear or see each other. However, these distances are still bigger than the scenario in which they are in the same room.
Smart distance refers to the situation where people intelligently adjust their distance based on various social contexts and preferences. For example, Figure 1(a) and Figure 1(b) show the social context where two persons adjust their distance by attitudes and by physical distances. As a matter of fact, distance adjusting appears in almost all social activities and working environments. A company brings people to work at the same location; however, it also allocates people to different rooms (Figure 1(c)).
Various technologies such as global networking, media
spaces [
Here we use a simple example to illustrate the concept. Suppose a team of five users is using Sametime Connect (Fig.2(a).) and Sametime Message (Fig.2(b).) as its communication interfaces among peers. Under usual situations, only Sametime Connect is used. If online chat is needed, Sametime Message must be used in order to perform the task. Thus, there are two choices of distance for peers of the team. Suppose user Yiming with address yiming@us.ibm.com has an online chat meeting scheduled with user Stephen with address levyzn@us.ibm.com at 5:00pm. Suppose user Yiming has a camera in his office that can detect whether he is in his office. Without smart distance, Yiming has to double click the item belonging to Stephen in Yiming’s Sametime Connect window to enable the online chat channel. If the system in Yiming’s office is designed such that it automatically turns on the Sametime Message window when the time is approaching 5:00pm and when the camera detects that Yiming is within his office, then the distance, a Sametime Connect display and a Sametime Message display, from Stephen to Yiming is a smart distance because it is selected automatically by the system based on contextual information. In the world of pervasive computing and global networking, there are various communication channels that can be provided by different pervasive devices; thus a huge number of different distances among peers can be selected. A system with smart distance ability will alleviate the user’s burden and help collaboration among users. The challenging tasks in designing such a system are how to detect the user’s intentions at any moment based on the sensing results of different pervasive devices and other contextual information and how to adaptively adjust distances among peers in favor of users’ intentions, various preferences, and the task at hand etc.
With ubiquitous connectivity on the horizon, and as more and more people work at the same time from different places, the issues of how to design the virtual organization and how to automatically adjust distances among people will become more and more important. The project “Smart Distance and WWWaware” is an effort along this line. Our goal is to build a multi-agent system called “Agent-Buddy” that can automatically detect different events associated with coworkers and can intelligently adjust awareness levels among co-workers. In this paper, we concentrate on the smart distance aspect of Agent-Buddy and study the influence of individual agent behaviors on the global team performance with respect to different kinds of tasks.
Software agents are studied from two complementary
perspectives. The first views software agents as entities with
different skills and knowledge within a larger community of
agents [
agent agent agent agent Fig. 3. The architecture of Agent-Buddy.
The Agent-Buddy approach is a combination of the above two approaches. Figure 3 shows the architecture of the AgentBuddy system. The goal of an agent in Agent-Buddy is to perceive events or status associated with one user and to selectively provide the perceived information to other users of the team. The Agent-Buddy system can be added to any CSCW system or virtual organization system to enhance the sense of working “together” concurrently and, at the same time, to keep the privacy of each user.
An agent in Agent-Buddy is a computational system that inhabits dynamic collaborative environments. It has knowledge about its own user and about the conventions of the working group. This knowledge can be used to guide its interactions with its responsible user and other agents of the group. The goal is to make the collaborative work easier and more efficient for members of the working group. There are two important features of the Agent-Buddy system. One is the event perception ability of each agent. The other is the automatic distance adjusting ability of the agent network. Each agent can perceive events with respect to its user based on signals perceived by all the devices within the user’s environment. In this paper, we concentrate on distance adjustment.
We have proposed a method that uses eigen-space and
eigen-pyramid to perceive events for agents [
To create a sense of group work, each agent has an “interface” to display events, through various means such as live video or live audio, associated with other users. Events associated with a user can be whether the user is logged on, how frequently the user is typing on the keyboard, what program the user is running, whether the user is entertaining himself by browsing the Internet, whether the user is working on the project, whether the user is on the phone and whom the user is talking to, whether the user is happy, sad or simply normal, whether the user has a visitor, and even whether the user needs a break because he is not being efficient at all, etc. However, an agent is not able to display all events of other users. There are two major concerns here. The first and the most important one is that the agent must communicate with agents of other users and must ask for permission to access events detected by those agents. It is up to other agents to decide what should be revealed to the asking agent. For example, events that intrude upon privacy cannot be accessed. For different asking agents, the criteria will be different. The second concern is that an agent should not display all the events of other users because it is usually unnecessary and impossible to do this within a single screen or through a multimodel interface. An agent must intelligently select events to display for the benefits of its user.
We use an awareness network to represent the awareness status provided by agents in the Agent-Buddy. An awareness network is a complete directed graph G=(V,D). Where V is the vertex set of G, and D is the edge set of G. Each element v ∈V corresponds to an agent in Agent-Buddy. For any two vertices vi and v j , there exist direct links dij and d ji (Figure 4). The link dij gives the distance from user i to user j, or in other words, it give the degree of difficulty for user j to perceive the activities of user i . It is a measurement of the amount of information about user i that is exposed to user j. The more the information is exposed, the smaller the value of dij . This value is selected by agent i by considering various factors. Similarly, d ji gives the distance from user j to user i.
i j d ij d ji
The values on the links of G are not constants; they keep updating at different times because of various factors such as the current tasks and the current events. Thus dij is a function of time. The awareness matrix, Ψ(τ ) = d11(τ ).....d1N (τ ) dN1(τ ).....dNN (τ ) gives the awareness status of the Agent-Buddy at time instant τ , where N is the total number of users in the system. Suppose that there are totally N d distances from one user to another, then the number of different statuses of the awareness network is: (N d ) n(n−1) . Where n(n −1) is the total number of directed edges in the awareness network G. At any moment, the state of the awareness network Ψ(τ ) can be one of the (N d ) n(n−1) states. The goal of the agents in Agent-Buddy is to automatically select, with consideration of various events and other factors, one state out of a huge number of potential candidate states at any moment such that the performance of the team is maximized or the performance is above a certain threshold. It is obvious that a central control mechanism will not work because of the complexity of the search space. Thus a distributed strategy is preferred.
For convenience, throughout the rest of this paper, we will use “agent i” to refer the “agent of user i”, and use “the distance from i to j” or “the distance from agent i to agent j” to refer to the “distance from user i to user j”.
We propose a physics-based framework for each agent to determine its distances to other agents, or in other words, the amount of information to expose to other users. This framework features dynamic models that incorporate various factors that must be considered.
Our idea comes from the elastic potential energy in physics. As is well known, if there is no force applied to a spring, the spring will be at its equilibrium position. However, if there are either compression or stretching forces applied to a spring, the spring will be deformed. The energy used to change the spring's displacement is stored in the coils as elastic potential energy. Most springs demonstrate a linear relationship between displacement from their natural positions and the applied forces and satisfy Hooke’s Law: F=-kx, where x is the displacement and k is a constant that measures the stiffness of a spring. The elastic potential energy is given by E = 12 kx2 . When Hooke’s Law is not satisfied, then the value of k will be a function of x and the potential energy can be given by: Ep = 0xk(t)tdt . Now, let us consider a physical system as shown in Figure 5. Figure 5(A) shows equilibrium positions of the springs. Figure 5(B) shows the situation when a horizontal massless flat plate is applied to this system. Each spring is connected to the plate. The potential energy of the springs system is given by:
E(x) = 0x−x1 k(t)tdt + + 0x−xn k(t)tdt . (2) Where x1, , xn are the natural positions of the springs and x is the final position of the plate and these springs. The x that minimizes E(x) is the equilibrium position for the system.
A Fig. 5. A physical system with many springs B
Now, let’s come back to the Agent-Buddy scenario and consider the task for agent i to determine the distance from i to j. In general, there exists an ideal amount of information that user i would like to be revealed to user j in a given situation. If agent i selects the distance that corresponds to the ideal situation, then there is no problem for user i at all. In most of the cases, however, agent i has to select a distance that is different from the ideal distance because of various reasons such as the special requirement of the current task etc. If the selected distance is different from the ideal distance, then there will be a tendency for user i to hope that the distance can come back to its ideal case. Let us imagine that there is a virtual spring for user i with a natural length that is equal to the ideal distance from user i to user j and that the selected distance is the actual length of the spring, Then when the selected distance is different from the ideal distance, there exists a virtual force that tries to pull or push the spring to its ideal length. The bigger the difference is, the stronger the force will be. We take the potential energy stored in this virtual spring as the measurement of the degree of anxiety or tension caused by the distance difference for user i. We call this energy the Backto-Ideal potential energy for user i. In the Agent-Buddy scenario, there are four kinds of factors to be considered by each agent. Each factor has a corresponding virtual spring. The goal of distance selection for each agent is to find a distance such that the total weighted Back-to-Ideal potential energies can be minimized.
The first factor is user i ’s current status and its associated ideal distance. Although one ideal distance might cover many situations, user i might need other ideal distances for some special events. For example, when user i is browsing the Web and having fun, he might not want user j to know about this activity although he might allow user j to monitor his activities under other situations. In any case, the ideal distances for user i under different situations might be different. We use diij,ev to represent the ideal distance for user i under the situation ev
The second factor is the requirement from the organization that uses Agent-Buddy. Because of the hierarchical structure of an organization and the complexity of relationships among the members, the awareness requirements to keep the organization functioning are different for different employees. For example, user i might be required to expose more information about himself to his manager than to his colleagues under other managers, and employees at the lowest level of the company may not be able to access any activity information of the CEO. We use d str to represent the ideal ij distance from i to j with respect to the organizational structure.
The third factor is user j’s request to user i on the amount of information user j would like to receive. Different users may want user i to expose different amount of information to them based on their own needs or preferences. We use dijj to represent the ideal distance from i to j with respect to user j’s request.
The fourth factor is the requirement for the current task. Different tasks require different awareness levels among team members. For example, a task that requires intensive discussions among team members such as brainstorming may require a higher level of awareness than a task that needs very few interactions. We use ditjq to represent the ideal distance from i to j with respect to the task tq .
Suppose that y is the distance selected by agent i, then the Back-to-Ideal potential energies of the above factors can be calculated as follows: (3) (4) δ 1 ( y) = 0y−diij,ev kiij,ev (x)xdx; δ 2 ( y) = 0y−disjtr kisjtr (x)xdx; δ ( y) = 0y−dijj kijj (x)xdx; (5) δ 4 ( y) = 0y−ditjq kitjq (x)xdx. (6)
Where kiij,ev , kisjtr , kijj , and kitjq give the stiffness of the virtual springs as a function of the offset x . When a virtual spring satisfies Hooke’s law, the corresponding stiff function is a constant and the Back-to-Ideal potential energy can be calculated easily. For example, if the first virtual spring satisfies Hooke’s law, then δ1( y) = 1 ki,ev × (diij,ev )2. 2 ij
In order to determine the final distance, agent i uses a weighted sum of Back-to-Ideal potential energies of the above factors as its objective function: δ ( y) = wiij ×δ1( y) + wisjtr ×δ 2 ( y) + wijj ×δ 3 ( y) + witjask ×δ 4 ( y), (7) where the weights encode user i’s preferences and determine agent i ’s behaviors in the collaborative environment. They are initially assigned to agent i by user i. The weight wiij specifies the degree of agent i ’s consideration on its user’s own need. The bigger the value of wiij , the more preferences agent i puts on its own user’s needs, thus the more selfish agent i is. The weight wisjtr gives the degree of how much user i emphasizes the organizational requirements. A bigger value of wstr ij corresponds to a better employee from the organizational point of view. The weight wijj gives the degree of the importance of user j in the mind of user i. A bigger value of wijj means that user i cares more about user j. It is also an indication of whether user i is cooperative with respect to user j. The (8) weight wtask gives the degree to which user i emphasizes a ij collaborative task. The higher the value of this weight is, the more collaborative agent i is.
Agent i will select the distance that minimizes δ ( y) as the distance from i to j: dij (τ ) = y* , such that ∀y,δ ( y* ) ≤ δ ( y). If all the virtual springs satisfy Hooke’s law, we have: y* = wiijkiij,ev diij,ev + wisjtrkisjtrdisjtr + wijjkijjdijj + witjaskkitjq ditjq
i i,ev + wisjtrkisjtr + wijjkijj + wtaskktq wijkij ij ij
In the above discussions, we assume that the stiffness function k(x) is a monotonous function with respect to a single-variable measurement of distance x . In most situations, however, distances are intrinsically multi-channel and may not be measured just by one single variable. For example, John and Mary are working in different places. Suppose that there are three ways for John to know Mary’s activities: (a) John can watch Mary’s activities only through a video camera installed at Mary’s office, the quality of the video can be adjusted; (b) John can listen to Mary’s activities only through an audio device installed at Mary’s office, the quality of the audio signal can be adjusted; and (c) John can watch and listen to Mary’s activities through both the above mentioned audio and video devices. It is easy for us to see that the distance from Mary to John for situation (c) is closer than that of situation (a) or that of situation (b), when the qualities of video are the same for all the situations and when the qualities of audio are the same for all the situations. However, it is a much more difficult job for us to compare the distances for situations of (a) and (b). We cannot really answer whether the distance for situation (a) is closer or further than that for situation (b), because they are coming from two different channels. Similarly, when the audio and video qualities are not constants, it is also difficult to compare two situations in situation (c). Thus, in this example scenario, distance d should be measured by two channels, d = d (a,v). Variable a refers to the quality of the audio signal. The higher the quality of the audio signal, the bigger the value of a . Variable v refers to the quality of the video signal. The higher the quality of the video signal, the bigger the value of v . A pair < a,v > defines a communication setting from Mary to John. All the different pairs of < a,v > determine the total possible communication settings from Mary to John. For any two different settings, distances may be comparable or may not be comparable; however, potential Back-to-Ideal energies can always be calculated because the difference in distances can always be obtained. Suppose we have two settings < a1,v1 > and < a2 ,v2 > . If a1 < a2 and v1 < v2 , then we have d (a1,v1) > d (a2,v2 ) . If a1 = a2 , then the function d (a1,v) , or d (a2 ,v) , is a monotonous decreasing function with respect to variable v . Similarly, if v1 = v2 , then the function d (a, v1) , or d (a,v2 ) , is a monotonous decreasing function with respect to variable a . On the other hand, if a1 > a2 and v1 < v2 , then we are not able to determine which distance is closer as there is no way to compare signals coming from two different channels. We, however, are able to calculate the Back-to-Ideal potential energy for any situations. Suppose d (a1,v1) is the ideal distance from Mary to John with respect to John and d (a2 ,v2 ) is the actual distance selected by Mary. If we imagine that there is a virtual spring within each channel, then the Back-toIdeal energy is the sum of frustrations caused by both audio a2 −a1 v2 −v1 and video and can be calculated by: ka (x)xdx + kv (x)xdx , 0 0 where ka and kv are stiffness functions for audio and video with respect to John. The differences in ka and kv reflect the relative importance of audio and video in John’s mind. In general, different channels have different stiffness functions for a given person. For the same channel, different people may have different stiffness functions.
In general, we need to first figure out how many channels a peer-to-peer communication can have, and then determine all the relevant stiffness functions involved. The channels should be selected in such a way that for a given channel, when the qualities of signals from all the other channels are fixed, the distance function should be a monotonous decreasing function with respect to the quality of the signal of the given channel. In the multi-channel scenario, a distance is no longer specified by a single variable as we did in Section V.B. Instead, a distance is specified by a set of variables that indicate the qualities of signals from all the channels. Suppose that there are totally z different channels. Let diij,ev ,cr be the ideal quality of the signal for channel r from user i to user j with respect to user i under the situation ev and let kiij,ev ,cr be the corresponding stiffness function. Then the Back-to-Ideal potential energy from i to j for user i with respect to signal quality setting s1, , sz is given by: δ1( s1, , sz ) = 0s1−diij,ev ,c1 kiij,ev,c1 (x)xdx + + sz−diij,ev ,cz ki,ev,cz (x)xdx . (9) 0 ij
δ 2 ( s1, , sz ) = 0s1−disjtr,c1 kisjtr,c1 (x)xdx + + 0sz−disjtr,cz kisjtr,cz (x)xdx ; (10) δ 3( s1, , sz ) = 0s1−dijj,c1 kijj,c1 (x)xdx + + 0sz−dijj,cz kijj,cz (x)xdx ; (11) δ 3 ( s1, , sz ) = 0s1−ditjq,c1 kitjq ,c1 (x)xdx + + 0sz−ditjq,cz kitjq,cz (x)xdx . (12)
Where the term δ 2 ( s1, , sz ) gives the Back-to-Ideal potential energy from i to j from an organizational structure point of view, when the signal quality setting is s1, , sz . The term δ 3 ( s1, , sz ) gives the Back-to-Ideal potential energy from i to j for user j , when the signal quality setting is s1, , sz . The term δ 4 ( s1, , sz ) gives the Back-to-Ideal potential energy from i to j from the point of view of the current task, when the signal quality setting is s1, , sz . The functions kisjtr,c1 , , kisjtr,cz , kijj,c1 , , kijj,cz , kitjq ,c1 , , and kitjq ,cz are the corresponding stiffness functions. Suppose that for channel cr ( r = 1, , or z ), there are totally zr ( r = 1, , or z ) different signal qualities. Then the total number of different signal quality settings is given by M = z1 × × zz . Each setting corresponds to a distance.
Similar to Section V.B, to determine the final distance, or the final choice of signal quality setting, agent i uses a weighted sum of Back-to-Ideal potential energies of the above factors as its objective function: δ ( s1, , sz ) = wiij ×δ 1 ( s1, , sz ) + wisjtr ×δ 2 ( s1, , sz ) + wijj ×δ 3 ( s1, , sz ) + witjask ×δ 4 ( s1, , sz ). (13)
Where the weights wiij , wisjtr , wijj , and witjask have the same meaning as those in Section V.B.
In most application situations, it is difficult to provide stiffness functions and to calculate the Back-to-Ideal potential energies. Furthermore, as illustrated in Section V.B, it might also be difficult to compare different distances given the multimodel nature of the Agent-Buddy. In order to avoid these difficulties, we propose a method that uses a set of Back-toIdeal energy difference vectors and matrices to guide agents in the selection of distances.
As analyzed in Section V.C, there are totally M different ways to expose one user’s status to another user. These M different ways correspond to M different distances d1 , …, dM among users. From a certain point of view, these distances encode the M different virtual walls among team members. Suppose that there are totally Q different events to be concerned with respect to users in Agent-Buddy.
The Back-to-Ideal potential energy matrix from i to j with respect to user i, Ηiij , is given by: Ηiij = h1i1 h1iM . Where huiv hQi1 hQiM gives the Back-to-Ideal potential energy when user i is at event u and agent i selected distance dv as the distance from i to j. If distance dv happens to be the ideal distance from i to j under event u with respect to agent i, then huiv = 0. In general, although user i might be at different states, only some special events might have different ideal distances. In most situations, user i’s ideal distance will be the same. Matrix Ηiij is available to agent i at the beginning and is specified by user i. The values of the elements of Ηiij encode the degrees of frustrations or tensions user i has for different selected distances under different events. Since the elements of the matrix provide the Back-to-Ideal potential energies with respect to user i, the calculation of δ 1 ( s1, , sM ) is avoided during the run time.
The Back-to-Ideal potential energy vector from i to j with respect to the organizational structure is given by: Hisjtr = (h1str , , hMstr ). Where hvstr gives the Back-to-Ideal potential energy with respect to the organization when agent i selects dv as the distances from i to j. If hvstr = 0, then dv is the ideal distance. The vector Hisjtr is provided by the organization to agent i at the beginning. Thus the calculation of δ 2 ( s1, , sz ) is avoided during the run time.
The Back-to-Ideal potential energy vector from i to j with respect to agent j is given by Hijj = (h1j , , hMj ). Where hvj gives the Back-to-Ideal potential energy from i to j with respect to agent j when agent i selects dv as the final distance. This vector encodes agent j’s preference of distances and is given by user j to agent j and is then passed by agent j to agent i. The calculation of δ 3 ( s1, , sz ) is thus avoided.
The Back-to-Ideal potential energy vector from i to j with respect to a given task tq is given by: Hitjq = (h1tq , , hMtq ). Where hvtq gives the Back-to-Ideal potential energy when agent i selects dv as the distance from i to j. The awareness requirements for a collaborative task might be given by the authority who assigns the task, or by the group conventions about the awareness level of the task, or by Agent-Buddy according to various experiences specified by users. In general, Agent-Buddy divides collaborative tasks into different categories according to the degree of awareness requirements for each member. It stores these tasks and the associated Backto-Ideal potential energy vectors in a common place such that each agent can retrieve the corresponding vector according to its role in the team. The potential energy vectors for all the tasks are available at the beginning, thus the calculation of δ 4 ( s1, , sz ) is avoided.
As discussed in the above section, the related Back-to-Ideal potential energies are all available for agent i. Thus, when a new collaboration task is assigned to user i or a new event is happening to user i, agent i will update the distances from its user to all the other related users.
Suppose that at time τ , user i is at the state of event u and the current collaboration task is tq , then the weighted Back-to
δ (dv ) = wiij × huiv + wisjtr × hvstr + wijj × hvj + witjask × hvtq . (14)
To select the best distance, agent i calculates the weighted Back-to-Ideal potential energies δ (d1), , δ (dM ) for all the distances d1, , dM and chooses the distance d with the minimum energy as the value of dij (τ ) , the distance from i to j at time τ . In other words, if δ (d ) ≤ δ (dv ) ( v = 1, , M ), then dij (τ ) = d .
At the beginning, all the agents within Agent-Buddy select their awareness distances to all the other agents according to the above method by assuming that there is no collaboration task. Thus, only the first three terms are involved in the calculation: δ (dv ) = wiij × huiv + wisjtr × hvstr + wijj × hvj . After Ψ(0) is determined, if there is no change in the status of any users and there is no new task, then the awareness status of Agent-Buddy will stay the same. This status will be updated whenever there are changes in events or tasks. When a change occurs, each related agent will update its distances to all the other agents according to the described method. The awareness status Ψ(τ ) of Agent-Buddy is adaptive to events and tasks. Each element dij (τ ) of Ψ(τ ) is an adaptive media wall in the virtual organization of Agent-Buddy. It is these virtual walls that keep the organization functioning and provide adaptive awareness to all the members of the team VII. THE INFLUENCE OF INDIVIDUAL BEHAVIOUR ON GLOBAL
PERFORMANCE
Here we study the influence of an individual agent’s behavior on the global team performance. There are many factors that can affect an individual agent’s behavior. For example, the Back-to-Ideal matrixes and Back-to-Ideal vectors influence an agent’s selection of distances. However, these factors encode the intrinsic properties of agents, the tasks at hand, and the organization. What we are interested in is how an agent’s personal properties, such as how it balances various preferences for itself, other agents, the task at hand and the organization, influence the outcomes of various kinds of global tasks. We hope that the empirical results along this line can provide some guidelines in the construction of virtual organizations.
We use the following virtual organization structure for our experiments. In Figure 6, each small circle represents an agent. If there is a line connecting two circles, then users represented by the two circles have a direct management relationship, where the one above is the manager of the one below. For example, user b is the manager of users e, f, and g. User i is the manager of user r, s, t, and u. In this organization, a is the CEO. We assume that users of this company work in distributed places, thus awareness plays a big role in the functioning of the company. Please note that this figure is not the structure of the Agent-Buddy for the organization. The structure of the Agent-Buddy is represented by a complete directed graph where circles of Figure 6 are vertexes of the graph and there are two directed links connecting each pair of vertexes. line manager is d95 . For example, distances from r to n, distances from r to q, distances from b to r are all equal to d95 . Figure 7 gives a subset of the ideal distance map of the organization.
b e f g h i
d l m
n a j c k u o p q r s t Fig. 6. Topological structure of the organization v w x y z
Suppose that there are 100 different distances d1, , d100 that can be used to provide awareness among co-workers of this company. Suppose that the smaller the index of the distance, the more the information is revealed to the receiving user. Thus, d1 provides the maximum awareness and d100 provides the minimum awareness.
The ideal distances from the organizational structure point of view are given as follows. The ideal distance from a user to his first line manager is d35 . For example, distances from b, c, and d to a, distances from e, f, and g to b, distances from x, y, and z to n, etc., are all equal to d35 . When we say that the distance from e to b is d35 , we mean that user b can check the activities of user e with the awareness degree given by d35 .
employee is d75 . For example, distances from a to b, c, and d, distances from c to i, distances from m to v and w, etc., are all equal to d75 . The ideal distance from a user to his second line manager is d55 . For example, the distance from e to a is d55 . The ideal distance from a user to his third line manager is d65 . For example, the ideal distance from q to a is d65 . The ideal distance from a second line manager to his second line employee is d91 . For example, the ideal distance from c to t is d91 . The ideal distance from a user to his third line manager is d65 . For example, the ideal distance from u to a is d65 . The ideal distance from a third line manager to his third line employee is d100 . For example, the ideal distance from a to u is d100 . This means that u has very little information on what user a is doing. The distances between any users that have the same first line manager is d50 . The distance between any users that do not share the same management chain or the same first e 50 o f p b
Suppose that for any agent, its ideal distances to any other agents are d77 and it hopes that any other agents can expose their activities at the awareness degree of distance d31 .
To make the discussion easier, we assume that the Back-toIdeal potential energy can be calculated according to Hooke’s law and that the stiffness function equals to constant 1 under all situations. Thus, the tension vectors and matrix can be easily obtained or calculated. For example, if the ideal distance is d50 and the selected distance is d65 , then the Back-to-Ideal potential energy can be calculated by: 1 ×1× (65 − 50)2 = 112.5 . 2
Our task is to evaluate how Agent-Buddy helps the productivity of a distributed collaborative work. This evaluation is based on how well the awareness provided to team members by agents of Agent-Buddy is, or in other words, how the various Back-to-Ideal potential energies or tensions are handled by those agents. The following formula is used to calculate the total Back-to-Ideal potential energies: ℜ(tq ) =
[K str ∆ str + Ktask ∆ task + K self ∆ self + Kothers ∆ others ]. (15) u∈Team(tq )
Where tq is the current task. Team(tq ) is the set of all the team members for the given task. ℜ(tq ) is the total weighted tension from all the agents of the Agent-Buddy related to this task. The higher the value of ℜ(tq ) is, the worse the performance. The contribution of each team member is the weighted sum of four factors. Weights K str , Ktask , K self , and Kothers give the sensitivity of the task with respect to awareness tensions in organizational structures, the current task, agents’ own expectations and other agents’ expectations respectively. They satisfy Kstr + Ktask + Kself + Kothers = 1 . ∆ str is the total structural tension from agent u to all the other agents related to the task. ∆ task is the total task tension from agent u to all the other related agents. ∆ self is the total self tension from u to other agents. ∆ others is the total tension with respect to other agents’ expectations from u to other agents. In practice, ∆ str , ∆ task , ∆ self , and ∆ others can be obtained from Hisjtr , Hitjq , Hiij , and Hijj as described in Section 5.3. Here we directly calculate the values of ∆ task , ∆ str, ∆ self , and ∆ others .
Since K str , Ktask , Kself , and Kothers give the properties of the task and wiij , wisjtr , wijj , and witjask determine agents’ behavior, we are able to study the influence of agents’ behaviors on the performance of Agent-Buddy by varying the above factors. In the following few experiments, we assume that agents b, f, q, i, and r are involved in the task.
Figure 8 shows the situation where agents’ concerns on structural needs can influence the system’s performance. The ideal distance for the task is d50 and it is neutral, which means that Kstr : Ktask : Kself : Kothers = 1 : 1 : 1 : 1 . For all the agents, the ratio of their behavior weights is given by wiij : wisjtr : wijj : wtask = 1 : wisjtr : 1 : 1 . Figure 8 shows how ℜ(tq ) and ij other Back-to-Ideal energies (tensions) are influenced when wisjtr changes from 0.1 to 10. We can notice that when agents put more weight on the organizational structure, the sum of the total structural Back-to-Ideal energies for all agents will decrease. The sum of the total “other” Back-to-Ideal energies for all agents will increase. This is because when agents emphasize structure more, they will put less weight on task awareness requirements and other agents’ awareness requirements. Thus, distance offsets with respect to these two factors will increase. It is interesting to note that the sum of the total “self” Back-to-Ideal energies for all agents will first decrease until the weight on structure equals to 3.7, and then the sum will increase. This is because when forces that pull distances toward the structural ideal directions become bigger and bigger, they also happen to pull distances towards directions of “other” distances from the global point of view. This situation will be changed when forces along “structure” direction are too big such that actual distances pass “self” distances and go to other directions. We can notice that the value of ℜ(tq ) will decrease until the weight on structure is around 1 and will increase after that. This tells us that for a neutral task, agents that extremely over-emphasize or overdeemphasize the organizational structure are not good. Thus, it is better to assign a neutral task to a group of agents that are also neutral.
Fig. 8. How agents’ concerns on structure influence the system performance. Figure 9 shows how agents’ concerns on other agents’ needs can influence the system performance. This time, the ratio of agents’ behavior weights is given by wiij : wisjtr : wijj : wtask = 1:1: wijj :1 , and wijj changes from 0.1 to 10.
ij We can notice that ℜ(tq ) is bigger when the weight is at 10 than it is when the weight is at 1. This tells us that sometimes extremely collaborative agents may not help team performance. It really depends on the nature of the task. The reason is that when agents are too concerned with other agents’ needs, the needs from other sources might be neglected. As a result, the performance as a whole might decline.
Figure 10 shows how the selfishness of an agent influences the system performance. The ratio of agents’ behavior weights is given by wiij : wisjtr : wijj : witjask = wiij : 1 : 1 : 1 , and wiij changes from 0.1 to 10. We can notice that the performance, ℜ(tq ) , reaches its minimum when wiij is around 1. Thus, for a neutral task, the more an agent emphasizes itself, the worse the performance.
Figure 11 shows how agents’ concerns on the awareness requirement influences the performance. The ratio of agents’ behavior weights is given by wiij : wisjtr : wijj : wtask = 1 : 1 : 1 : witjask , ij We can notice that the and wtask changes from 0.1 to 10.
ij performance is best when witjask is around 1. When agents over emphasize the awareness requirement of the task, the performance declines rather than enhances. This is because the property of the task itself is neutral, thus, a departure from its own requirement may not influence the success of the task with big impact. However, it does influence other factors. Thus, the combined results will reduce the performance of the team. situation that wisjtr changes from 0.1 to 10 while other weights all equal to 1. We can notice that the best performance occurs when the factors are around position 1. This further illustrates that when agents’ behaviors match the properties of the task, the performance of the system will be high.
Now, we change the property of the task in another test such that Kstr : Ktask : Kself : Kothers = 5 : 7 : 5 : 9 . The task is no longer neutral. Thus, neutral behaviors of agents will not generate the best performance. This is shown in Figure 13. We can also notice that when corresponding weights pass the value of 1, the Back-to-Ideal energies related to structure and the Back-toIdeal energies related to agents themselves increase much faster than the other cases. This is because the task does not emphasize these factors. Thus, if agents over-emphasize them, the performance of the team will decrease. In general, the performance of the team depends on various factors and can be very complex. Based on our extensive experiments, we find that in most situations, a better match of agents’ behaviors and task properties tends to provide a better team performance.
Figure 14 shows how the value of ℜ(tq ) will be influenced when we change agents’ behavior weights wiij and the task property
Kself . .
Here the ratio for agents is: wiij : wisjtr : wijj : wtask = wiij : 1 : 1 : 1 , where wiij changes from 0.1 to ij 10.
The ratio
for Kself : Kstr : Kothers : Ktask = Kself : 1 : 1 : 1 .
the task is: The term Kself changes from 0.1 to 2.7. The line shows that when the properties of the task and the properties of the agents match, the system obtains its best performance.
Many researchers have addressed the issue of multiagent
collaboration within a multi-user environment [
In this paper, we propose the concept of distance and smart distance in a distributed collaborative environment. We illustrate an Agent Mediated Collaborative system - the AgentBuddy system that can create a sense of group presence and at the same time preserve the privacy of each user. We define the multiagent awareness network to represent the awareness situations among team members in a virtual organization or in a CSCW scenario. A virtual spring is used to model the awareness degree among team members. Each agent makes decisions by considering multiple factors. The goal of the multiagent team is to minimize the global awareness frustrations with respect to different kinds of tasks. Empirical studies have been conducted to analyze the individual agent behavior on the global performance.
With ubiquitous connectivity on the horizon, collaborative computing will become one of the major applications in the evolution of computing and communication. The goal of our research is to dynamically adjust the “distance” among people in a collaborative environment - breaking the isolation, providing group awareness, and at the same time, keeping the privacy. It is our belief that researches in multi-user and multiagent aspects of virtual organizations such as an awareness network will become more and more important. Our vision is that in the future, WWW will not only be the knowledge pool of human society, but also be the digital world where people can meet and sense each other through various pervasive devices, virtual reality techniques, and peer-to-peer networking.
user user user user user knowledge pool user user user
User
User User adjustable distance
User User
User Break the isolation! Fig.15.. In the future, Internet will be a communication channel that provides WWW peer-to-peer awareness and interaction through various pervasive devices and virtual reality techniques.
The authors would like to thank Barbara Grosz, Luke Hunsberger, Sanmay Das, Dave Sullivan, Tanara Babaian, Jill Nickerson, Wheeler Ruml, and Tim Rauenbusch of Harvard University, Catalina Danis, and Alison Lee of IBM Research, Candace Sidner of MERL, Clay Shirky of The Accelerator Group for fruitful discussions.