University of Nebraska at Omaha

DigitalCommons@UNO

Computer Science Faculty Publications Department of Computer Science

1-2006

Topologies of agents interactions in knowledge intensive multi-

agentsystems for networked information services
Qiuming Zhu
University of Nebraska at Omaha, [email protected]

Follow this and additional works at: https://digitalcommons.unomaha.edu/compscifacpub

Part of the Computer Sciences Commons

Please take our feedback survey at: https://unomaha.az1.qualtrics.com/jfe/form/
SV_8cchtFmpDyGfBLE

Recommended Citation
Zhu, Qiuming, "Topologies of agents interactions in knowledge intensive multi-agentsystems for
networked information services" (2006). Computer Science Faculty Publications. 28.
https://digitalcommons.unomaha.edu/compscifacpub/28

This Article is brought to you for free and open access by

the Department of Computer Science at
DigitalCommons@UNO. It has been accepted for
inclusion in Computer Science Faculty Publications by an
authorized administrator of DigitalCommons@UNO. For
more information, please contact
[email protected].
1 58
2 59
3 60
4 61
5 62
6 63
7 64
8 Topologies of agents interactions in knowledge intensive multi-agent 65
9 66
10
systems for networked information services 67
11 68
12 Qiuming Zhu * 69
13 70
Department of Computer Science, University of Nebraska at Omaha, 6001 Dodge St., Omaha, NE 68182, USA
14 71
15 72
16 73
17 74
18 Abstract 75
19 76
Agents in a multi-agent system (mAS) could interact and cooperate in many different ways. The topology of agent interaction determines how
20 77
the agents control and communicate with each other, what are the control and communication capabilities of each agent and the whole system, and
21 78
how efficient the control and communications are. In consequence, the topology affects the agents’ ability to share knowledge, integrate
22 79
knowledge, and make efficient use of knowledge in MAS. This paper presents an overview of four major MAS topologic models, assesses their
23 80
advantages and disadvantages in terms of agent autonomy, adaptation, scalability, and efficiency of cooperation. Some insights into the
24 applicability for each of the topologies to different environment and domain specific applications are explored. A design example of the 81
25 topological models to an information service management application is attempted to illustrate the practical merits of each topology. 82
26 83
27 84
28 85
29 1. Introduction agent’s tasks, or may not contain that a closely looking 86
30 supervisor—like a real estate agency, as long as every agent 87
31 Software agents, one of the most exciting new develop- operates in compliance with the agency operating protocol (e.g. 88
32 ments in computer software technology, can be used to quickly following work ethics, paying fees on time). The underlying 89
33 and easily build integrated enterprise systems. The software agent architecture must support sophisticated reasoning, 90
34 agents, like people, can possess different levels of competence learning, planning, and knowledge representation of the 91
35 at performing a particular task. The idea of using multiple individual agent or the agencies. A general understanding of 92
36 software agents that communicate and cooperate with each a MAS is that: (i) each agent has a partial capability to solve a 93
37 other to solve complicated problems in various complicated problem, (ii) there is not necessary a global system control, (iii) 94
38 personal and enterprise computing application domains on our data and knowledge for solving the problem are decentralized, 95
39 behalf is intuitively appealing. One significant benefit of multi- and (iv) computations carried out among the agent are 96
40 agent systems (MASs) is their scalability. Since they are asynchronous [13]. 97
41 inherently modular, it is easier to add new agents to a multi- MAS contain extremely high-level of software abstractions. 98
42 agent system than it is to add new capabilities to a monolithic Programming an agent-based system is primarily a matter of 99
43 system. specifying agent behavior. In MAS, the agents need to work 100
44 Agents in a MAS can have different functionalities and collectively so that, as a group, their behavior solves the overall 101
45 behaviors. For example, agents can be categorized as self- problem without disruption, conflict, and glitches. When a task 102
46 governing agents, brokered agents, monitored agents, mediated is assigned, the agents are likely in needs to find the other 103
47 agents, etc. Each individual agent can be crafted to be an expert agents to collaborate with. Such a task is easy if they know 104
48 in solving a specific problem or performing a particular task. A exactly which agents to contact and at which location. 105
49 collection of software agents that communicate and cooperate However, a static distribution of agents is very unlikely to 106
50 with each other is called an agency. An agency may have a exist. For dynamic multi-agent systems, agents need to know 107
51 manager that closely supervise and arrange the individual how and where to find the other agents [16]. The dynamic 108
52 nature of agent distribution motivates this research to look at 109
53 the topological models of MAS and study how these models 110
54 * Tel.: C1 402 554 3685; fax: C1 402 554 3400. facilitate or hurdle the agent collaborations. 111
55 E-mail address: [email protected] Software developers and system designers use high-level 112
56 abstractions in building complex MAS. To manage 113
57 the complexity, MAS abstraction must focus on the important 114
115 and essential properties of a problem and hide the incidental amplified when propagated through the agent system. More- 172
116 components of that problem. An agent interaction topology over, some agents may become disconnected from the rest of 173
117 provides a simple way of managing the complexity because a the system by temporary or permanent disability of these agents 174
118 topology is essentially a high-level abstraction about the or their communication channel, resulting in incomplete/ 175
119 interactions of the functional components in a complex system inconsistent system states. How should we represent individual 176
120 such as the MAS. The topology of agent interaction also helps agents acting in such an uncertain environment, and more 177
121 to define (or facilitates the definitions of) the communication importantly, how can we predict how the MAS as a whole will 178
122 protocol and the interface among the agents of MAS. evolve as the result of uncertain inter-agent interactions? 179
123 It is understood that in a complex system, each agent only Properly structured topology plays a critical role to address 180
124 needs to interact with a limited number of agents, most likely the above problems in MAS systems. The topology determines 181
125 the agents in its vicinity. Agents in MAS can be organized and how the agents interact with human and with each other, what 182
126 controlled in many different ways. For example, agents could are the relations among the agents, and how data and 183
127 be entitled as equal right citizens. That is, every agent has the knowledge are shared and communicated among the agents 184
128 same status and control and access right to other agents and [18,20]. The topology would also affect the functionality, 185
129 their shared resources. In this case, each agent would have the capacity, and underlying computation mechanisms of the agent 186
130 same capability of solving a given problem [3]. Who does what assembly. To date, there have been relatively few implemen- 187
131 purely depends on who is available at the moment. The benefit tations of complex agent-based systems. The difficulty of 188
132 of this model is that the system is highly fault tolerant—leave determining what agent system topology to employ partly 189
133 one or two agents out of the cycle, the job still gets done as limited the more spacious spreading of MAS in real world 190
134 usual. Moreover, the agents in this model have the maximum applications. A proper topology leads to desirable collective 191
135 degree of autonomy. They volunteer their service by behavior in large and complex MAS. Therefore, MAS research 192
136 themselves upon a request of service or inbound object/ needs an insight on how different architectural topologies of an 193
137 situation/environment changes. One other choice is a hier- agent assembly function differently to the effects toward agent 194
138 archical model in which agents are grouped/labeled with adaptation, control, collaboration, and learning [12, 195
139 different classes/status in terms of the functionality or assigned Grefenstett296]. 196
140 rights [28]. These agents are often under a centralized or an In this paper, we first present an overview of four major 197
141 upper level control. Some supervisory agent in the system may MAS topology models. They are (1) a Web-like topology 198
142 be identified. This organizational model has the advantage of where agents are connected (and communicated) as nodes in a 199
143 operational efficiency and configuration flexibility [Sohata94]. complete graph; (2) a Star-like topology where several agents 200
144 Software agents are suitable for use in a wide variety of are connected with, and communicate through, a controller/- 201
145 applications. However, agents can have different ways of inter- mediator; (3) a Grid-like topology where each agent is only 202
146 connections and interactions. Each of the interaction schemes connected (and communicated) with its neighboring agents, 203
147 is appropriate for use in implementing certain kinds of thought the access to other agents or resource not in the 204
148 applications. Developers must carefully analyze system neighborhood could be done through the neighboring agents; 205
149 requirements to determine if the selected agent interaction and (4) a hierarchical collective agent network (HCAN) 206
150 scheme is an appropriate implementation mechanism. The topology, that combines some of the features of previous 207
151 study of the structural and cooperative topology is necessary models. We assess the advantages and disadvantages of these 208
152 for construction of complex systems involving multiple agents models in terms of agent autonomy, adaptation, scalability, and 209
153 and mechanisms for coordination of independent agents’ efficiency of cooperation. An example of the application of the 210
154 behaviors toward a common goal. MAS can be considered of fourth model for application in information service is 211
155 containing the following four dimensions [11]: (1) Agent presented. 212
156 granularity (coarse vs. fine); (2) Heterogeneity of agent The paper is organized as the following. Section 2 discusses 213
157 knowledge (redundant vs. specialized); (3) Methods of the four major MAS agent cooperation topologies. Section 3 214
158 distributing control (benevolent vs. competitive, team vs. assesses these four topologies in terms of a set of criteria 215
159 hierarchical, static vs. shifting roles); and (4) Communication selected. Section 4 presents an analysis of the fourth topologies 216
160 possibilities (blackboard vs. messages, low-level vs. high- with respect to different MAS application domains, and points 217
161 level, content). The MAS designers must consider the some insights on the applicability of each topology to certain 218
162 capabilities of each individual agent and how multiple agents applications. Section 5 presents an exemplar design of using 219
163 can work together—the architecture and protocol issues. There each of the topologies for an information service system 220
164 are many ways and views in the study of multi-agent system application. Section 6 contains conclusion remarks. 221
165 architecture and protocol. In this paper the architecture and 222
166 protocol issues are explored from the topological point of view. 2. Taxonomy 223
167 Development of multi-agent system (MAS) applications is 224
168 often complicated by the fact that agents operate in a dynamic, Several research communities have modeled distributed 225
169 uncertain world. Uncertainty may stem from noisy external computing by studying communication and coordination 226
170 data, inexact reasoning such as abduction, and actions by mechanisms among autonomous software entities, or agents. 227
171 individual agents. Uncertainty can be compounded and Agent-based computing focuses on the interaction mechanisms 228
229 among agents, which permit a rich set of coordinated activities. 286
230 Effective models of interaction require the following basic 287
231 capabilities: 288
232 289
233 (1) A transport mechanism to convey messages in an 290
234 asynchronous fashion; 291
235 (2) An interaction protocol, defining the available types of 292
236 communications and their semantics; 293
237 (3) A content language providing the base for composition of 294
238 requests and their interpretation; and 295
239 (4) An agreed-upon set of shared vocabulary and meaning of 296
240 concepts (often called on ontology). Fig. 1. Web-like topology of agent interaction. 297
241 298
The degree to which different agents play out distinct roles In the Web-like topology, the collection of distributed
242 299
is certainly an important issue in MAS. The taxonomy agents acts as equal members of the community. In this
243 300
presented in this paper is organized along the most important topology, all of the agents have the same internal structure as
244 301
aspects of agents: degree of heterogeneity and degree of well as operation goals, domain knowledge, and possible
245 302
communication for interaction and knowledge sharing. The action choices. They also have the same procedure for selecting
246 303
taxonomy is based on the common understanding that: (1) among their actions. The only differences among agents may
247 304
agents are ubiquitous, (2) agents have designated roles, reside be their sensory inputs and the actual actions they take: they
248 305
at designated place, perform designated tasks for a designated may be situated differently in the world or in different
249 306
person/controller, and (3) agents can be acting by their own environmental settings. Although the agents have identical
250 307
(once deployed) or agents can be acting under coordination of capabilities and decision procedures, they may have limited
251 308
other agents. information about each other’s internal state and sensory
252 309
The topology of multi-agent cooperation can be classified inputs. Thus they may not be able to predict each other’s
253 310
according to multiple criteria. In this paper, we use the actions.
254 311
255 following three criteria to characterize the cooperation: The Web-like topology can also be formed in virtual when 312
256
the MAS employs an agent-activation scheme called request- 313
(1) The ways of activation, supervision, and communication and-service protocol, a blackboard kind of communication and
257 314
between the agents [18], i.e. how the agents invocate each task activation approach. In the request-and-service protocol,
258 315
other, requesting service from each other, and retrieve/pass every agent in the MAS can response to a call issued by one of
259 316
data to each other; the agent and perform the task requested, and could be called
260 317
(2) The dependencies of the agents [19], i.e. whether they by other agents to perform specific tasks. That makes the agents
261 318
function complementary to complete a task, i.e. each seemed all connected directly.
262 319
functioning on the same course or differently aspects of a In the Web-like topology, the agents are empowered as
263 320
course, and equal-right citizens in a MAS society. Every agent receives the
264 321
(3) The ways of sharing data, knowledge and other resources, same command and request, share the same data and resources,
265 322
including considerations of at what level they share the and act at the same level (though functioning differently in
266 323
data and knowledge to complete a given task [30]. terms of the problem to be solved). Each agent can call any
267 324
268 other agents, and be called by any other agents. The General 325
In this section we study four basic MAS topological Magic’s MAS model is a representative example of this kind of
269 326
structures: (1) a Web-like topology, (2) a Star-like topology, topology [34]. General Magic models MAS as an electronic
270 327
(3) a Grid-like topology, and (4) a Hierarchical Collective marketplace that lets providers and consumers of goods and
271 328
Agent Network (HCAN) topology. Note that this study is not services find one another and transact business. This market-
272 329
about the physical links between the agents. Our concern is on place is modeled as a network of computers supporting a
273 330
the functional links (and interactions) among the agents collection of places that offer services to mobile agents. All
274 331
enabled either by physical links or by virtual communication
275 agents have the same capability to travel from one place to 332
channels. The four MAS topologies of our study are described
276 another, to meet other agents which allows them to call one 333
in the following.
277 another agent’s procedures, to create connections to allow an 334
278 agent to communicate with another agent in a different place, 335
279 2.1. Web-like topology and to have authority to indicate the real-world individual or 336
280 organization that the agent represents. Note that in Web-like 337
281 A Web-like topology is featured with a uniform inter- topologies, agents can perform their service by themselves 338
282 connection of the agents in a cooperative environment. That is, autonomously upon a request of service (ROS) or inbound 339
283 every agent node can have directly interaction with all other objects or situation/environment changes. 340
284 agent nodes. Usually, these interactive agent nodes form a A number of variations to the Web-like model exist. For 341
285 complete graph, as shown in Fig. 1. example, the agents are organized in groups (subsets) and 342
343 agents in each subset are fully connected in the Cougaar MAS with descriptions of the capabilities of other agents, and then 400
344 architecture. The Cougaar architecture supports a distributed delegate the tasks to qualified/available agents [8]. Thus, it is 401
345 plan, similar to a partitioned blackboard, which is inter- not generally required that a requester (user or agent) know the 402
346 connected but not replicated across the agent society [9]. This identities, locations, or number of other agents involved in 403
347 means that information is shared among only the interested satisfying a request. Facilitators are not viewed as centralized 404
348 parties. This simple concept, combined with some proven controllers, however, but rather as coordinators, as they draw 405
349 concepts of locality of reference, minimizes the communi- upon knowledge and advice from several different, potentially 406
350 cation requirements and makes possible a managed agent distributed, sources to guide their delegation choices. This 407
351 network required of large-scale distributed systems. scheme makes it possible for software services to be provided 408
352 through the cooperative efforts of distributed collections of 409
353 autonomous agents. 410
2.2. Star-like topology
354 In a distributed agent framework of Star-like topology, a 411
355 dynamic community of agents, where multiple agents 412
Unlike in Web-like topology where agents can be contribute services to the community, is often conceptualized.
356 413
cooperative in their own all together by some implicit When external services or information are required from a
357 414
agreement or activation protocol, there may be actions that given agent, instead of calling a known subroutine or asking a
358 415
require explicit coordination for successful execution. In a star- specific agent to perform a task, the agent submits a high-level
359 416
like topology, the activities of the agents are coordinated or expression describing the needs and attributes of the request to
360 417
administered by some supervisory (or facilitator) agents a specialized facilitator agent. The facilitator agent will make
361 418
designated in the assembly. Only agents that have connections decisions about which agents are available and capable of
362 419
363
built and specified to the coordinator can interact with each handling sub-parts of the request, and will manage all agent 420
364
other. That is, the agents are more under control and stipulation interactions required to handle the complex query. One 421
365
than those in the Web-like topology. In this topology, advantage of this quasi-distributed agent architecture is that 422
366
functional invocation and data communication is often it allows the construction of MAS that are more flexible and 423
367
brokered through connections to one or more facilitating adaptable than the fully distributed object frameworks such as 424
368
agents. The facilitator is responsible for matching requests those in the Web-like topology. Individual agents can be 425
369
from users to agents, with descriptions of the capabilities of the dynamically added to the community easily, extending the 426
370
agents in its possession. A structural diagram of such topology functionality that the agent community can provide as a whole. 427
371
is shown in Fig. 2, where the center nodes in dark color denote The agent system of Star-like topology is also able to adapt to 428
372
the coordinators. available resources in a way that hard-coded distributed objects 429
373 Most agent architectures contain specialized agents that are systems cannot. 430
374 suited for specific operations within the application domain and One of the important issues to consider when designing a 431
375 environment. Often sophisticated systems of application were multi-agent system is whether the different agents will be 432
376 decomposed into modules, each of which was then transformed benevolent or competitive. Even if they have different goals, 433
377 into an agent or multi-agents. These agents then are divided the agents can be benevolent if they are willing to help each 434
378 into different groups. Agents in each group are capable of other achieve their respective goals [15]. On the other hand, the 435
379 performing a specific kind of tasks. In this configuration, the agents may be selfish and only consider their own goals when 436
380 agents may not communicate with each other directly. A acting. In the extreme, the agents may be involved in a zero- 437
381 supervisor, controller, or mediator is then needed to distribute sum situation so that they must actively oppose other agents’ 438
382 and coordinate the tasks. Examples of such control agents goals in order to achieve their own. The Star-like topology is 439
383 include (1) the SRI’s OAA facilitator [24]; (2) the CMU’s more empowered to solve these kinds of goal and action 440
384 RETSINA Matchmaker [32]; and (3) the Infosleuth’s Broker conflicts among the group of agents. 441
385 [26]. 442
386 In SRI’s Open Agent Architecture (OAA), the facilitators 2.3. Grid-like topology 443
387 are responsible for matching requests from users and agents, 444
388 In a grid-like topology, each agent cooperates with a group 445
389 (an agency) of agents in its neighborhood (in terms of 446
390 functional connections) that is a subset of agents in the 447
391 assembly (or community). Each agent has direct connections 448
392 (in terms of cooperation behavior) to the agents in its 449
393 neighborhood group (logically, not necessary physically or 450
394 geographically). Each group may be administered by a 451
395 supervisor/facilitator designated. Interaction to agents not 452
396 residing in the neighborhood must pass through the facilitators 453
397 of the neighborhoods. Such interaction may pass multiple 454
398 agents in cascade. The designation of facilitator may be 455
399 Fig. 2. Star-like topology of agent cooperation. changed dynamically in terms of the efficiency of interaction it 456
457 2.4. HCAN topology 514
458 515
459 A fourth topology, named a hierarchical collective MAS 516
460 model, is presented in this section. The hierarchical collective 517
461 agent network (HCAN) topology of agent cooperation is shown 518
462 by diagram in Fig. 4. Main properties of the HC topology are 519
463 (1) Agents are grouped in layers, (2) the layers are organized in 520
464 hierarchy, (3) agents in each layer are not connected, (4) agents 521
465 between layers are fully connected, and (5) the control and 522
466 coordinate of the agent at each layer are through the agents at 523
467 the higher level. 524
Fig. 3. Grid-like topology of agent cooperation.
468 In the HCAN, agents at the lower level (the data managing 525
469 enables. Fig. 3 shows a diagrammatic illustration of this module) interface directly to individual sensor/information 526
470 topology, where the nodes in dark color denote facilitators resources. These agents act in a distributive fashion to process 527
471 under current designation. conceptual queries, filter retrieved information using simple 528
472 Simply described, a grid-like topology is an environment proposition logics, and extract useful information as instructed 529
473 consisting of areas. Areas are required to have exactly one local by upper-level (the reasoning or user interface modules) 530
474 area coordinator, which is an agent that acts as a facilitator for agents. The agents at the upper levels coordinate the activities 531
475 other agents within its area. Agents can be identified as being of the agents at the lower levels using a centralized goal-driven 532
476 inside an area if they have registered with the area’s local 533

F
control strategy. They issue conceptual queries, perform data
477 coordinator. Agents will use the services of local area integration and knowledge extraction, and make cross- 534

O
478 coordinators to access other agents in the system. Agents can 535
reference of the information retrieved. The coordinate agents
479 advertise services and find out about other agents’ services by 536
at these levels will apply certain data analysis models and

O
480 means of agent registry or yellow page servers. Agents 537
employ reasoning-integration technique to fuse information
481 requiring data sharing with other agents can join virtual 538
reported by retrieval agents at the lower levels. Special human-
482 environments called cooperation domains, which are supported
PR 539
system interfacing agents will provide continual support for
483 by cooperation domain server agents. 540
interactions between user and the systems, and provide
484 541
The agents in Grid-like topology form a more federated intelligent and dynamic information summarization, annota-
485 542
agents society. It has relatively low communication and tion, and presentation based on the user-originated inputs and
486 543
computational requirements, meaning that there are virtually queries.
D

487 544
no constraints on the system size. The simplicity of agent The major functionalities and design tradeoffs of the HCAN
488 545
interactions also makes it amenable to quantitative mathemat- topology are as follows. The HCAN topology is flexible in
TE

489 546
ical analysis. Each group of agents has a meta-agent that serves terms of the ability in which communities of agents can be
490 547
as the agent/task manager, which decomposes a task and assembled, and the flexibility with which services can be added
491 548
distributes it to the individual functional agents or other agent at runtime and brought into use without requiring changes to
492 549
managers. Example of MAS in the grid-like topology can be the other part of the agent assembly. A unified set of concepts,
EC

493 550
494
seen at the Object Manager Group (OMG)’s Model [33]. This declarations and interfaces that are consistent across all 551
495
model is composed of agents (i.e. components) and agencies services in the framework, and the role played by the agents 552
496 (i.e. places) as entities that collaborate using general patterns at different levels are defined. The HCAN topology strikes a 553
R

497 and policies of interaction. Under this model, agents are balance between the centralized control and distributed 554
498 characterized by their capabilities (e.g. inference, planning, computation by allowing distributive agent operations within 555
R

499 and so on), type of interactions (e.g. synchronous, asynchro- layers of the hierarchy and enforcing centralized control 556
500 nous), and mobility (e.g. static, movable with or without state). between the layers of the hierarchy, thus eases the coordination 557
O

501 Agencies, on the other hand, support concurrent agent and control needed to manage interactions between agents. 558
502 execution, security and agent mobility, among others. 559
In many systems, hierarchically organized collections of
C

503 560
504 planning agents that are committed to one particular planning 561
problem are deployed. For example, in MPA- Multi-agent
N

505 562
506 Planning Architecture of SRI [35], the activities of these agents 563
are coordinated by meta-PAs (PAs that control other PAs) with
U

507 564
508 specialized knowledge about strategies for division of labor, 565
509 conflict resolution, and (in the future) plan merging. Each 566
510 meta-PA is responsible for coordinating activities among its 567
511 collection of PAs and other planning clusters. 568
512 569
513 Fig. 4. Hierarchical collective topology of agent cooperation. 570
571 Table 1 (2) Cooperative. Agents in a MAS should be specially 628
Features of four major MAS topology equipped with the ability to work with other agents to
572 629
573 Web Star Grid HCAN achieve a common goal. They must behave effectively at 630
574 Center controller /mediator? No Yes Partly Partial both self-organizing and delegating states, effective under 631
575 Agents all at equal level? Yes No No No coordination and negotiation, and conscious of conflict 632
576 One to all interaction? Yes No No No resolution. 633
577 Complete communication link? Yes No Partly Partial (3) Trustful. The agents must be reliable when exerting their 634
Local/global distinction? No No Yes Yes
578 autonomy in performing the tasks designated by human. 635
Automatic service response? Yes No Partly Partial
579 They must perform the tasks and complete the tasks in the 636
580 quality and time as the human instructed. 637
581 (4) Flexible. Agents in MAS should be flexible in terms of 638
582 The rationale behind the HCAN topology is again the system reconfiguration and task delegation. Agents should 639
583 concept of shared and distributed intelligence. It is not a good be able to join and participate the cooperation community 640
584 idea to develop agents with capability of doing everything. at any time, i.e. dynamic inhabitation. Configuration 641
585 Agent must be task-specific for doing something, and for doing flexibility leads to scalability that is also critical to MAS 642
586 some small things really well. That is, agents are specialists on operating in dynamic environment. 643
587 special tasks. For example, it is not necessary to require an (5) Adaptive. Agents should have a certain level of ability to 644
588 agent to possess all the perception, action, and reasoning selectively sense and act/re-act to the environmental 645
589 components, which are necessary for being autonomous and situation changes, and should be readily/easily transplan- 646
590 adaptive. Rather, it can be an agent system in which there are table to different environmental applications. 647
591 agents responsible for perception, agents responsible for (6) Interactive. Most agents are required to communicate and 648
592 action, agent responsible for reasoning, and agents responsible interoperate efficiently with humans, other systems, and 649
593 for learning and augment the knowledge of the other agents or information sources. Agents in MAS must be especially 650
594 accumulate and store the knowledge to a place that are capable of dealing with the complexity issues of resource 651
595 accessible by all the agents. Where the perception agents feed sharing, distribution, and deadlock breaking. 652
596 the reasoning agents, the reasoning agents feed the action (7) Reactive. The ability to learn and improve the functionality 653
597 agents, and the learning agents feed both the reasoning and with experience is a very desirable feature of agents. 654
598 action agents, etc. Thus, the functionality of an agent must Agents able to dynamically adapt to and learn from the 655
599 always be limited to a specific domain, on a specific task. That environment will have better capability to adapt to 656
600 is, based on this observation and understanding the MAS situation/environment changes. 657
601 comes into play. 658
602 659
603 2.5. Summary 3.1. Web-like topology 660
604 661
605 Table 1 Summarizes the structure characteristics of the Both advantages and disadvantages of the Web-like 662
606 above four MAS topology. topology are associated with its indiscriminative behavior of 663
607 agent activation. The Web-like MAS topology facilitates 664
608 3. Analyses parallelism and entitles redundancy. While parallelism is 665
609 achieved by assigning different tasks or abilities to different 666
610 In this section we explore the advantages and disadvantages agents, robustness is a benefit of multi-agent systems that have 667
611 of the topologic models of the above in terms of their effects to redundant agents. If control and responsibilities are sufficiently 668
612 agent autonomy, adaptation, communication, learning, and shared among different agents, the system can tolerate failures 669
613 efficiency of cooperation. The topology should facilitate the by one or more of the agents. Domains that must degrade 670
614 intensive knowledge embedding, accumulation, and incorpor- gracefully are in particular need of this feature of MAS: if a 671
615 ation for MAS. A multi-agent system is dynamic in nature, single entity -processor or agent- controls everything, then the 672
616 meaning that agents can be added to it or removed from it from entire system could crash if there is a single failure. 673
617 time to time. Thus, an agent system topology must also One question often asked of this kind of MAS is that in such 674
618 facilitate the dynamic property of agents. The study here a closely coupled relation among agents—agent network, can 675
619 focuses on how the specific topology boosts or attenuates the agents be really equal members of a society? Or, is this 676
620 major agent features and functionalities required by MAS, especially good for the joint functionality of a MAS? The 677
621 based on a set of agent properties defined as the following: answer may depend on what application domain the agent 678
622 system works in. Although multi-agent systems are often 679
623 (1) Autonomous. It is known that agents, whether in a MAS or described as being intrinsically more robust than a single agent 680
624 stand-alone, should be proactive, goal directed and act on by virtue of redundancy, fault tolerance is not a natural 681
625 their own (self-starting behavior) or perform tasks on some byproduct of duplication but only emerges through careful 682
626 user’s behalf. Effectiveness of goal achieving is one design. A complex MAS cannot always be created through 683
627 important property of agents. cloning a group of single agents designed for the same task. 684
685 There has to be some awareness, either on the part of the agents 3.3. Grid-like topology 742
686 or the system designer, of the role that other members will play 743
687 in completing the task. Unless the global task is somehow The grid-like topology makes a tradeoff between increasing 744
688 partitioned among the agents, they will either interfere with the number of agents that can interact directly with each other 745
689 each other or converge on a sub-optimal division of labor. and retain control of monitoring of agent activities in a 746
690 Thus, the reason why a complete-graph kind of topology is not reasonable range. The approach is suitable for MAS designed 747
691 necessary, and probably undesirable, is that the global to operate in a well-defined global environment and objectives. 748
692 interaction with all agents in a domain or application The topology entitles the relative merits of model-free and 749
693 environment is likely not necessary. Moreover, the design of model-based methods. Consider the facilitating of local or 750
694 that kind of global interaction system is too complex to deal networked configuration of the MAS as another criterion, the 751
695 with. The functional structure of individual agent in Web-like grid topology is advantages than the other topologies of MAS. 752
696 topology is also most complex among the topologies because The locally interacted agents in Grid-like topology may 753
697 the agent there needs to know how to communicate with the demonstrate complex group behavior advantages over the fully 754
698 others, while in other topologies the communication can be connected agent assembly. When agents have similar goals, 755
699 they can be organized into a team. Each agent then plays a 756
handled by the facilitator or broker agent.
700 separate role within the team. With such a benevolent team of 757
701 agents, one must provide some method for assigning different 758
702 3.2. Star-like topology agents to different roles. This assignment might be obvious if 759
703 the agents are very specific and can each only do one thing. 760
704 An advantage of star-like topology is its loosely enforced However in some domains, the agents are flexible enough to 761
705 control and coordination. Though control and coordination interchange roles. 762
706 limits the boundary of cooperation the agents can reach, it is 763
707 desirable when efficiency of cooperation is a main issue that 764
708 needs to be ensured. The star-like topology is suitable for the 3.4. HCAN topology 765
709 environment and applications where the MAS is to act as a 766
710 central planner, that involves team negotiation and needs The HCAN topology makes a tradeoff between distributive 767
711 awareness of what each agent knows and does. It also possesses and centralized control of multiple gent systems. The collective 768
712 functional suitability and self-consciousness—each agent is nature of the agents in the HCAN paradigm overcomes some of 769
713 dissimilar in functionality, the dissimilarity determines and these difficulties, for example, relieving the burden of data- 770
714 distributes tasks. The use of facilitators in OAA offers both exchanges between fellow agents by limiting agent communi- 771
715 cation to vertical layers of the assembly only. The collective 772
advantages and weaknesses with respect to scalability and fault
716 nature of agent relation in the hierarchical architecture 773
tolerance [6]. For example, on the plus side, the grouping of a
717 simplifies the functional design of the agent interactions and 774
facilitator with a collection of client agents provides a natural
718 enhances the security and efficiency of the information 775
building block from which to construct larger systems. On the
719 processing. 776
minus side, there is the potential for a facilitator to become a
720 Basically, the HCAN is desirable when the MAS is required 777
communication bottleneck, or a critical point of failure.
721 to have the following functionalities. 778
In Star-like topology, the control agent focuses on the
722 779
interaction mechanisms among agents, which permits a rich set (1) A flexible software architecture for accommodating
723 780
of coordinated activities. Effective models of interaction system augmentation and evolutions;
724 781
require some basic capabilities: (1) a transport mechanism to (2) A powerful representation schema for accommodating
725 782
convey messages in an asynchronous fashion, (2) an interaction heterogeneous forms of information;
726 783
protocol, defining the available types of communications and (3) A diverse interface for various input resources, output
727 784
their semantics, (3) a content language providing the base for formats, and human interactions;
728 785
729
composition of requests and their interpretation, and (4) an (4) An ability of reasoning on incomplete and inconsistent 786
730
agreed-upon set of shared vocabulary and meaning of concepts information, and extracting useful knowledge from the 787
731
(often called on ontology). Some MAS use game theoretic data of heterogeneous resources; 788
732 model for multi-agents cooperation and rely on the assumption (5) An ability of incorporating real-time dynamics of the 789
733 that all agents are fully rational. In general, for a set of agents to information resources into the system anytime during the 790
734 cooperate, there is a need for a shared ontology among them. It operation, and promptly adjusting the reasoning mechan- 791
735 is more critical to have a shared ontology for agents to inter- isms; 792
736 operate without passing through a facilitator. (6) An ability of summarizing and refining knowledge 793
737 Another advantage of mediated topology is that it is easy to extracted, and distinguishing mission and time critical 794
738 define a system in terms of agent-mediated processes. The knowledge from insignificant and redundant ones; 795
739 moderated multi-agent systems are particularly well suited to (7) A capability of supplying meaningful and accurate 796
740 process and workflow automation, electronic commerce, explanations, both qualitatively and quantitatively, of the 797
741 distributed problem solving, Internet applications. automated system actions; and 798
799 Table 2 the diverse sets of MAS applications. It is noted that most of the 856
Assessment of the topologic models agent research and development up to date are in the area of
800 857
801 Web Star Grid HCAN agent modeling and agent building tools. Wide spreading true 858
802 Autonomy 5 1 3 4 applications are still lacking. Over hundred agent construction 859
803 Cooperative 2 5 3 4 toolkits, development environment, or component libraries can 860
804 Trustful 1 5 5 5 be returned from a simple search on Internet. Chauhan and 861
805 Flexible 5 5 5 4 Baker, 1998’s JAFMAS supports directed (point to point) 862
Adaptive 2 5 5 5
806
Interactive 3 1 3 5 communication as well as subject based broadcast communi- 863
807 Reactive 2 5 3 5 cations [5]. Ciancarini et al [7] introduced PageSpace as a 864
808 referential architecture for designing interactive multi-agent 865
809 applications, using variants of the coordination language Linda 866
(8) A capability of providing adequate control and scrutinizing
810 to guide their interactions. Several kinds of agents live in the 867
of the system operations under the environmental
811 PageSpace: user interface agents, personal home agents, agents 868
constrains of the given situation.
812 that implement applications, and agents that interoperate with 869
813 There is a need for mechanisms for advertising, finding, legacy systems. Suzuki et al. [31] proposed ‘self-migrating 870
814 fusing, using, presenting, managing, and updating agent threads’ as a new cluster-computing paradigm for multi-agent 871
815 services and information in most MAS applications. To applications, which can be viewed as the interactions among 872
816 address these issues, the notion of middle agents was proposed autonomous computing entities, each having its own objec- 873
817 [11,22,23]. Middle agents are entities to which other agents tives, behavior, and local information in a synthetic world. 874
818 advertise their capabilities, and which are neither requesters Self-migrating threads have both navigational autonomy of 875
819 nor providers from the standpoint of the transaction under mobile agents and fine computation granularity of threads. In 876
820 consideration. The advantage of middle agents is that they ZEUS [25], coordination is supported through use of 877
821 allow MAS to operate robustly when confronted with agent 878
conversation classes that agents utilize to manage their
822 appearance, disappearance, and mobility. There are several 879
interactions with other agents during problem solving. The
823 types of agents that fall under the definition of middle agents. 880
conversation classes implement rule based automata models,
824 Note that these types of agents, which are described below, are 881
similar in spirit to the way co ordination behavior is managed
825 defined so vaguely that sometimes it is difficult to make a clear 882
in ZEUS.
826 differentiation between them. 883
Multi-agent systems (MASs) provide for the modeling of
827 884
practical systems in the fields of communications, flexible
828 , Facilitators. Agents to which other agents surrender 885
manufacturing, and air-traffic management [4,27]. Some of the
829 their autonomy in exchange for the facilitator’s services. 886
previous work in multi-agent system development concen-
830 Facilitators can coordinate agents’ activities and can 887
trated on domain-independent frameworks, standard protocol
831 satisfy requests on behalf of their subordinated agents. 888
definitions, some handling of uncertainty and utility, and
832 , Mediators. Agents that exploit encoded knowledge to 889
extensive models of collaboration [16]. However, there lacks
833 create services for a higher level of applications. 890
methods for solid decision-theoretic model of agents learning,
834 , Brokers. Agents that receive requests and perform 891
adaptation, control and collaboration. Arai et al presented a
835 actions using services from other agents in conjunction 892
reinforcement learning approach known as Profit-sharing that
836 with their own resources. 893
allows agents to learn effective behaviors with in dynamic and
837 , Matchmakers and yellow pages. Agents that assist 894
838 service requesters to find service provider agents based multi-agent environments [1]. The increased prevalence of 895
839 on advertised capabilities. agents raises numerous practical considerations. Three of these 896
840 , Blackboards: Repository agents that receive and hold are (1) adaptability to unforeseen conditions, (2) behavioral 897
841 requests for other agents to process. assurance, and (3) timeliness of agent responses [2,14]. Two 898
842 questions are always asked about any type of technology. (1) 899
843 The HCAN provides a proper balance on the need of What advantages does it offer over the alternatives? And (2) In 900
844 centralized and distributed middle agents for the control and what circumstances is it useful? The same questions apply to 901
845 coordination of the multi-agents in the complex system. the study of topologies of MAS. The evolution of Multi-Agent 902
846 The assessments of the four major topologies are Systems and the growing interest in multi-agent development 903
847 summarized in Table 2. We give a rating of 1–5 to each of platforms have leaded to some interesting tools for agent 904
848 the performance measurements for each topology, where a software developers. Although, some platforms are grounded 905
849 rating of 1 is the lowest and 5 is the highest. The assignments on well-known models, platforms for development of agents 906
850 are somehow subjective. are widely heterogeneous globally. Questions remaining: What 907
851 topology of agent interaction is good for what kind of 908
852 4. Applications applications? 909
853 We first take a look at some examples to see the diversity of 910
854 After comparing the four basic topological structures and MAS applications and what kind of cooperation topology is 911
855 their pros and cons, we can now relate the major topologies to needed for each of the applications. 912
913 1. An electronic commerce application might have buyer While a peer-to-peer processing application has significant 970
914 agents, seller agents, stocking agents, database manage- advantages over the client-server approach in these 971
915 ment agents, email agents, etc. A loan approval application applications, agents in these systems must be highly 972
916 ties together branch banks, the main bank, loan under- autonomous meanwhile trustful. 973
917 writing companies, and credit reporting companies, and 974
918 automates much of the loan approval process. All of these Table 3 categorizes the major applications of MAS, with 975
919 agents involve distributed computation or communication respect to the features of the application domain, specific 976
920 between components, need to communicate with each problems deal with, and features of each type of the 977
921 other, and must have the capability of working together to applications related to agent characteristics. 978
922 achieve a common set of goals. Multi-facets of consider- It would be desirable to have a statistics on the variations of 979
923 ations must be made with respect to the differences in MAS applications and the major system topology employed in 980
924 performance efficiency and competency when choose each of the applications. There are two main factors that make 981
925 proper topology for the agent system in these applications. it difficult to enumerate the application systems with respect to 982
926 2. Data fusion and mining applications that reason about the the topologic types of the agent interactions. One is the limited 983
927 messages or objects received over a network require multi- resource available for the real world MAS applications, 984
928 agents organized in sequences of work-flow and coordi- especially lacking the application systems with significant 985
929 nation, e.g. network interfacing agent, information search- influence to the field. The second is that in many real 986
930 ing agent, recording agents, inference agents, reporting applications, there is no clear cut on which topology the 987
931 generation agents, etc. The same situation applied to agents in the system apply. More often the applications have a 988
932 e-collaboration and e-learning applications. Agent system mixture of the interaction topologies among the interactions of 989
933 in these applications must balance the distributiveness and the agents in the applications. Instead, we thus turned to a look 990
934 centralized control. at the MAS development/construction tools (toolkits, 991
935 3. Automation applications for example in plant and process languages, libraries) to find the correspondences of the 992
936 automation, workflow management, robotics including topology enabled/allowed by these systems/tools. We have 993
937 Unmanned Autonomous vehicles (UAV), etc. requires the evaluated 26 commercial and 39 academic MAS products 994
938 agent to be capable of operating without much user input or and/or development packages/toolkits. Tables 4 and 5 995
939 intervention. An embedded factory controller might consist summarize the systems. It is found that no any of the above 996
940 of a user interface agent, a database interface agent, a topology is in a dominating position in either domain. 997
941 machine tool interface agent, and a process monitoring and However, two observations are worth to mention. One is that 998
942 control agent. All of these agents could run concurrently on while the Star-like topology was seen in 28% of academic 999
943 the same processor or could be easily distributed across systems, there is no (0%) any commercial system adopting this 1000
944 multiple processors. scheme. The other is that the grid-like topology is the most 1001
945 4. There are applications that require significant communi- popular one in both the commercial (23%) and academic (36%) 1002
946 cations between components for sensing or monitoring of systems. Note that quite an amount of systems also possesses 1003
947 the environment, making decisions and performing auton- the property as a mixture of both grid-like and star-like 1004
948 omous operations. Since the agents in these applications topology. If we consider this mixture topology together with 1005
949 need to have the ability to reason (i.e. draw inferences), the grid-like ones, then a majority in both academic and 1006
950 they can easily perform sequences of complex operations commercial systems is present. 1007
951 based on messages they receive, their own internal beliefs, It is not our intention to collect and summarize all published 1008
952 and their overall goals and objectives. For example, email MAS application systems that have been built or reported. 1009
953 and instant messaging system that uses software agents to Therefore our discussion will be focused on the categories of 1010
954 implement the mail client. The system is designed to ensure applications, without referring to specific products or product 1011
955 that messages remain private. Privacy is assured systems. We thus present an extensive, but not exhaustive, list 1012
956 because messages never reside on any server device. of work in the field. Despite the youth of the field, space does 1013
957 1014
Table 3
958 MAS systems with respect to application domains 1015
959 1016
Domain of application Features of the application Type of agents in need Suitable topology Complexity of interaction
960 1017
961 Information service Mixture of distributive and centralized Diverse Grid or HCAN Low 1018
Web search Distributive uniform Web-like Low
962 1019
Planning and Scheduling Centralized, semi-distributive Heterogeneous Star-or Net-like Mild
963 Process control (manufacture Semi-distributive, mixture of distribu- Diverse Grid or HCAN High 1020
964 assembly, air traffic) tive and centralized 1021
965 Reasoning and decision making Mixture of distributive and centralized Mixtures HCAN high 1022
966 Data fusion and mining Centralized Mixtures Star or grid or mild 1023
967 HCAN
1024
Simulation Mixture of distributive and centralized Diverse kinds Star or grid High
968 E-commerce Peer-to-peer uniform Web-like low 1025
969 1026
1027 Table 4 will provide continual support of interactions between IMS and 1084
Commercial MAS development/construction products: total 26 the agents. The hypothetic information service management
1028 1085
1029 Topology type Number of systems Percentage system must accommodate the following agent assemblies. 1086
1030 W 5 19 The information service broker agent. The information 1087
1031 S 0 0 service broker assembly contains three agents: Publish Service 1088
1032 G 6 23 Agent (PSA), Subscribe Service Agent (SSA), and Query 1089
1033 G/S 6 23
Service Agent (QSA). These agents interface directly to the 1090
H 0 0
1034 information clients to manage the Pub/Sub/Query Services. 1091
Other 9 35
1035 The agent functions can be defined as the following. 1092
1036 Star topology: there seemed to be no instances of a star topology in the 1093
commercial realm. Because of the size of deployment (load/volume) in a
1037
commercial realm vs. academia, that would explain why a star would be
(1) The PSA possesses the functions of (a) Processing the 1094
1038 deployable in academia, but not in a commercial arena. G/S: the combination of requests of permission for publish from the publisher (a 1095
1039 G/S meant that there were options within the framework to allow for either a client), through interactions to I&A (Identification and 1096
1040 single entity to perform the controlling function of agents or to distribute that Authentication) agent. (b) Creating a publisher sequence 1097
1041 control in a more grid-like pattern. H topology: actually found an instance of 1098
with the client once permission is granted. (c) Receiving
the Hierarchical in the academic arena. It was described in the product info
1042 and transmitting the metadata and payload provided by the 1099
almost exactly what your paper describes. Other: many commercial products
1043 that would probably be classified in the academia world as grid-like, are publisher under a publication request, thereby creating an 1100
1044 actually classified as other in commercial because that called themselves a tool IO (Information Object) in the IOR (IO Repository). (d) 1101
1045 to build tools for marketing purposes. In that sense it could be called a Providing a universally unique identifier (UUID), created 1102
1046 particular ‘type of topology’ but the product information was somewhat 1103
confusing. by the IOR agent, back to the publisher for future
1047 reference. 1104
1048 not permit exhaustive coverage. Instead, the work mentioned is 1105
(2) The Subscribe Service Agent (SSA) will possesses the
1049 intended to illustrate the techniques that exist to deal with the 1106
functions of: (a) Processing the subscriber’s requests for
1050 issues that arise in the various multi-agent scenarios. 1107
permission to subscribe, through interaction to I&A. (b)
1051 1108
Processing the subscription predicate (subscriber metadata
1052 1109
5. Example constraint) that the platform applies over the MDR
1053 1110
(Metadata Repository) of newly published IOs to
1054 1111
In the following we present an example design of determine delivery. (c) Notifying the subscriber of
1055 1112
application of MAS with the four topologies studied in this available IOs, generally done thru a client-defined call-
1056 1113
1057 paper. We know that software agents provide a powerful new back. 1114
1058 method for implementing the next-generation information (3) The Query Service Agent (QSA) possesses the functions of 1115
1059 systems. In the example multi-agent system described below, (a) Processing Query client’s requests of permission to 1116
1060 agents are designed to perform information gathering, query, through interaction to I&A. (b) Informing the Query 1117
1061 categorization, and distribution according to specific needs of client to submit a query request containing a query 1118
1062 users. Special human-system interfaces built in these agents metadata constraint to the platform, once permission is 1119
1063 Table 5
granted. (c) Returning a set of partial result IOs based on 1120
1064 Academic MAS development/construction products: total 39 the access control policy established for the particular 1121
1065 client. 1122
Topology type Number of systems Percentage
1066 1123
W 2 5 The information management expedition agents. The
1067 1124
S 11 28 information management expedition assembly contains the
1068 G 14 36 1125
1069 G/S 8 21 agents for IOR, MDR and I&A management. These agents 1126
1070 H 1 3 function as the following. 1127
1071 Other 3 8 1128
1072
(1) The IOR agent manages and performs the archiving and 1129
Star topology: there seemed to be no instances of a star topology in the
1073 commercial realm. Because of the size of deployment (load/volume) in a organization of published IOs for later retrieval by 1130
1074 commercial realm vs. academia, that would explain why a star would be subscribe and query. The IOR agent is capable of handling 1131
deployable in academia, but not in a commercial arena. G/S: the combination of a throughput of millions of IOs and hundreds of IO types at
1075 1132
G/S meant that there were options within the framework to allow for either a a time.
1076 single entity to perform the controlling function of agents or to distribute that 1133
1077 control in a more grid-like pattern. H topology: actually found an instance of (2) The MDR agent manages and supplies clients with 1134
1078 the Hierarchical in the academic arena. It was described in the product info information about available IO types to which the client 1135
1079 almost exactly what your paper describes. Other: many commercial products has access. The MDR contains all schemas and other data 1136
that would probably be classified in the academia world as grid-like, are for approved IO types and versions within the platform.
1080 1137
actually classified as other in commercial because that called themselves a tool
1081 to build tools for marketing purposes. In that sense it could be called a
(3) The I&A agent associates and ensures a unique identifier 1138
1082 particular ‘type of topology’ but the product information was somewhat with each client/administrator, issues and verifies the 1139
1083 confusing. authenticator and credentials based on open standards to 1140
1141 the maximum extent with little or no modification of client 5.1. Web-like topological implementation 1198
1142 code. 1199
1143 Note that in this example, agents are classified with different 1200
1144 The information system control agents. The information functionalities. However, the interactions among the agents are 1201
1145 system control assembly contains the account manage agent nevertheless organized in a Web-like topology. This means that 1202
1146 (AMA), access control agent (ACA), and persistence every agent in the system is capable of communicating and 1203
1147 adaptation agent (PAA). These agents function as the interacting with each other. The interaction diagram is shown 1204
1148 following. in Fig. 5. 1205
1149 The major advantage of the Web-like topological 1206
1150 (1) The Account manage agent (AMA) is responsible for implementation of the system is that versatile agent functions 1207
1151 creation of accounts that include issuance of authenti- can be built and incorporated into the system and interaction 1208
1152 cators and credentials; modification of accounts to broadly overall the system. The major problems with this 1209
1153 include disabling accounts, and changing privilege levels implementation are that (1) it is somehow hard to solve the data 1210
1154 via re-issuance of credentials; deletion of accounts. inconsistency problem once it happens among the agents, for 1211
1155 (2) The Access control agent (ACA) is responsible for example, for subscribe service, publishing service, and the IOR 1212
1156 granting access to IOs and system resources to maintenance; (2) it is incapable of generating and disseminat- 1213
1157 authorized clients and administrators. An access control ing user-tailored information under dynamical changes of the 1214
1158 mechanism is enforced by the agent that only allows for situation because adaptation to such a change requires complex 1215
1159 the dissemination and receipt of IOs in compliance coordination of goal and functional specification changes 1216
1160 with the platform access control policy. among a number of agents, and (3) the control structure of each 1217

F
1161 (3) The Persistence adaptation agent (PAA) has the agent is rather complicated because of the heterogeneity of the 1218

O
1162 capability to manage the lifecycle of information within agent modules in the system. Since there is no central 1219
1163 the platform, ensures interoperability and the system’s controller or mediator, all the control functions among the 1220
diverse of agents must be built into each individual agent. We

O
1164 survival of several generations of clients without 1221
1165 degraded service over time. While the IMS (Information do not recommend such implementation for the supposed 1222
1166 manage Staff) is solely responsible for removing information service management system.
PR 1223
1167 information objects from the information space, the 1224
1168 PAA provides the means to accomplish this in 5.2. Star-like topological implementation 1225
1169 accordance to policy established. 1226
1170 A Star-like topological implementation of the hypothetic 1227
D

1171 Thus, the entire exemplar information service manage- information service management system has the agent 1228
1172 ment system consists of nine agent modules. In the interaction diagram as shown in Fig. 6. 1229
TE

1173 following, we illustrate the simulative implementation of In this topological implementation, one extra agent in 1230
1174 the information service management agent system in the addition to the nine required agent modules is employed in the 1231
1175 four topologies, respectively. system architecture. The additional agent, named Agent 1232
1176 1233
EC

1177 1234
1178 Account MDR 1235
1179 maintenance maintenance 1236
1180 Agent Agent 1237
R

1181 1238
Subscribe Query
1182
Service
MDR Service
1239
R

1183 Agent Agent 1240

1184 1241
O

1185 1242
1186 1243
C

1187 1244
1188 1245
N

1189 I&A 1246

maintenance IOR Persistence
1190 1247
Agent adaptation
U

1191 Agent 1248

1192 1249
Access Publish
1193 control IOR Service 1250
1194 Agent maintenance Agent 1251
1195 Agent 1252
1196 1253
1197 Fig. 5. Web-like topology of MAS for information service management. 1254
1255 1312
Account MDR
1256 maintenance maintenance 1313
1257 Agent Agent 1314
1258 1315
Subscribe Query
1259 MDR 1316
Service Service
1260 Agent Agent 1317
1261 1318
1262 Agents 1319
1263 controller 1320
IOR IOR
1264 and 1321
1265 coordinator 1322
1266 I&A Persistence 1323
maintenance adaptation
1267 1324
Agent Agent
1268 1325
1269 Publish 1326
Access
1270 control IOR Service 1327
1271 Agent maintenance Agent 1328
Agent
1272 1329
1273 1330
1274 Fig. 6. Star-like topology of MAS for information service management. 1331
1275 1332
1276 1333
Controller and Coordinator, is located in the centralized the dynamics of the environment and requirement changes of
1277 1334
position among the agents. It has two-way direction connection the system. Here the role of PAA is also different from the
1278 1335
to all the agent modules, while the information service agents Controller and Coordinator agent in the Star-like topology in
1279 1336
do not directly interact with each other. The advantages of this the way that the PAA does not take the charge of coordinate the
1280 1337
scheme are that (1) it is easy to solve the data inconsistency execution of the interacting agents. The agents in the system all
1281 1338
problem, and guarantee the right information retrieval and have certain level of autonomy in terms of performing their
1282 1339
delivery, and (2) it is possible to have additional agents with designated tasks. The agent interaction diagram is shown in
1283 1340
versatile functions, such as data fusion and mining, added to Fig. 7.
1284 1341
the service, assuming the agent controller and coordinator Major advantage of this Grid-like topological implemen-
1285 1342
maintains properly an agent registry that allows for dynamical tation is that the functionality of the individual agent can be
1286 1343
addition or deletion of agents in the assembly. Disadvantages optimally conducted because the agents are connected in the
1287 1344
of the implementation are (1) it would be less efficient to way that only those necessary interactions are permitted.
1288 1345
execute the information retrieval and delivery functions However, this implementation makes it hard to adjust and
1289 1346
because each of these function requires activation of at least modify the agent configuration, thus limits the versatility of
1290 1347
two agents, the coordinate agent and the subscribe or publish functions can be incorporated in to the system. The control
1291 1348
agents, and (2) while the control structure of the information structure of overall system is also relatively complicated. This
1292 1349
service agents will be less complex because each of them only implementation thus is also not in our recommendation.
1293 1350
need to interact with the controller, the control structure of the
1294 1351
coordinator agent will be relatively complicated. This
1295 5.4. HCAN implementation for information service 1352
topological implementation would be a choice if the security
1296 management 1353
and reliability is the main concern and the efficiency (rapid
1297 1354
performance of the information service functions) is not a
1298 The design of HCAN architecture and algorithms expedite 1355
major issue.
1299 the integration of publishing, subscribing, and query services in 1356
1300 a heterogeneous information space. The system is organized in 1357
1301 5.3. Grid-like topological implementation three agent layers, as shown in Fig. 8: (1) a information service 1358
1302 broker layer at the lower level of the hierarchy; (2) a 1359
1303 In a Grid-like topological implementation, we place the information expedition layer at the middle level of the 1360
1304 Persistence Adaptation Agent (PAA) at the center of the hierarchy; and (3) a system control layer at the top level of 1361
1305 assembly and the other agent modules surrounding it. the hierarchy [21]. The functionalities of these layers are 1362
1306 However, it differs from the Star-like topology in the way described in the following. 1363
1307 that the other agents all have interactions with their The information service broker layer contains subscribe, 1364
1308 neighboring agents, in addition to the interactions with the publish, and query agents to interact with the information 1365
1309 PAA. The PAA is chosen sit in the center because its service clients and networked information sources, respect- 1366
1310 functionality may be need to all the other agents, for example, ively. These agents detect and collect data, perform key word, 1367
1311 adjusting the agent functional parameters according to string, or context extractions from the data feeds, and submit 1368
 Q. Zhu / Advanced Engineering Informatics xx (xxxx) 1–15

1369 MDR 1426

1370 maintenance 1427
Subscribe Agent
1371 Query 1428
Service
1372 Agent Service 1429
Agent
1373 1430
1374 MDR MDR 1431
1375 1432
1376 1433
Account Persistence Access
1377 maintenance adaptation control 1434
1378 Agent Agent Agent 1435
1379 1436
IOR IOR
1380 1437
1381 1438
1382 1439
1383 I&A Publish 1440
1384 maintenance IOR Service 1441
Agent maintenance Agent
1385 1442
Agent
1386 1443
1387 1444
1388 Fig. 7. Grid-like topology of MAS for information service management. 1445
1389 1446
1390 1447
filtered reports to the upper level agents for information interacting with human operators of this information service
1391 1448
package and delivery. system.
1392 1449
The information expedition layer accommodates three The advantages of HCAN topological implementation are
1393 1450
information contents level management agents to perform (1) the agents are better under control of appropriate agents that
1394 1451
coordination tasks for information object repository mainten- enables efficiency of each agent’s performance meanwhile
1395 1452
ance, metadata repository maintenance, and information source ensures the reliability of the operations, and (2) the MAS
1396 1453
1397
identification and authentication. structure is flexible to add additional agents with versatile 1454
1398
The system control layer contains agents to support the functions, such as data fusion and mining. Since only agents 1455
1399
information service level management tasks, such as the client between layers are connected via heterogeneous links and are 1456
1400 account maintenance and access control, and persistence interactive, each agent is relatively independent. This makes 1457
1401 adaptation that performs tasks to adapt the system to the additions of agents and modifications of the agent 1458
1402 environmental variation or requirement changes. The user functionalities simple. Major disadvantage of the implemen- 1459
1403 interface and system management functions are also performed tation is that it requires a little more deliberated planning, 1460
1404 by the management agents at this layer that in charge of design, and understanding of the interaction logics of 1461
1405 1462
1406 1463
1407 Account Account Persistence 1464
Manage Manage adaptation
1408 1465
Agent Agent Agent
1409 1466
1410 1467
1411 1468
1412 MDR MDR 1469
IOR MDR I&A
1413 Manage Manage Manage 1470
1414 Agent Agent Agent 1471
1415 1472
1416 IOR IOR 1473
1417 Query 1474
Publish Subscribe
1418 Service Service Service 1475
1419 Agent Agent Agent 1476
1420 1477
1421 1478
1422 Info. Info. Info. 1479
1423 Client 1 Client 2 Client N 1480
1424 1481
1425 Fig. 8. HCAN topology for information service management. 1482
1483 the agents distributed on different layers. Overall, the HCAN References 1540
1484 topological implementation is our recommendation for the 1541
1485 intended information service management system. [1] Arai T, Sycara K, Payne T. Experience-based reinforcement learning to 1542
1486 acquire effective behavior in a multi-agent domain Proceedings of the 6th 1543
pacific rim international conference on artificial intelligence 2000 pp.
1487 1544
125–35.
1488 6. Conclusions [2] Barbuceanu M, Fox MS. Cool: a language for describing coordination in 1545
1489 multi agent systems Proceedings of the first international conference on 1546
1490 The agent-based system developments have emerged from multi-agent systems (ICMAS-95).: AAAI press; 1995 pp.17–24. 1547
1491 their primarily functional diversities to the stages that raise the [3] Bradshaw JM. An introduction to software agents. In: Bradshaw JM, 1548
1492 necessity of managing the system complexity. Building editor. Software agents. Menlo Park, California: AAAI Press; 1997. 1549
p. 3–46.
1493 reliable, maintainable, extensible, and re-usable MASs that 1550
[4] Bradshaw JM, Dutfield S, Benoit P, Woolley JD. Toward an industrial-
1494 conform to their specifications requires modeling techniques strength open agent architecture. In: Bradshaw JM, editor. Software 1551
1495 that support abstraction, structuring, and modularity. The most agents. Menlo Park, California: AAAI Press; 1997. p. 375–418. 1552
1496 widespread methodologies developed for the conventional [5] Chauhan D, Baker A. JAFMAS: a multiagent application development 1553
1497 software systems are various object-oriented approaches. They system Proceedings of autonomous agents 98. New York: ACM Press; 1554
1498 have achieved a considerable degree of maturity and are 1998 pp. 100ndash;7. 1555
[6] Cheyer A, Martin D. The open agent architecture. J Auton Agents Multi-
1499 supported by a large community of software developers. The 1556
Agent Syst 2001;4(1):143–8.
1500 system architecture of object-oriented systems is based on the [7] Ciancarini P, Tolksdorf R, Vitali F, Rossi D, Knoche A. Coordinating 1557
1501 notion of objects, which encapsulate state information as data multiagent applications on the WWW: a reference architecture. IEEE 1558
1502 values and have associated behaviors defined by interfaces Trans Softw Eng 1998. 1559
1503 describing how to use the state information. Object oriented [8] Cohen PR, Cheyer AJ, Wang M, Baeg SC. An open agent architecture. In: 1560
1504 formal approach address almost all the steps in the process of Oren Etzioni, editor. Proceedings of the AAAI spring symposium series 1561
on software agents; 1994, 1994. p. 1–8.
1505 designing and implementing a software system, providing a 1562
[9] http://www.cougaar.org.
1506 uniform paradigm across different system scales and [10] Decker K. Distributed problem solving: a survey. IEEE Trans Syst Man 1563
1507 implementation languages. However, there are additional Cybern 1987;17(5):729–40. 1564
1508 issues related to the development and implementation of [11] Decker K, Sycara K, Williamson M. Middle-agents for the internet 1565
1509 multi-agent systems that need to take serious care of. Proceedings of the international joint conferences on artificial intelligence 1566
1510 The implementation of multi-agent systems involves a great (IJCAI-97) 1997. 1567
number of problems with respect to the components, protocols, [12] Ferguson IA, Karakoulas GJ. Multiagent learning and adaptation in an
1511 1568
information filtering market Adaptation, coevolution and learning in
1512 interactions, and schemes. In particular it is often hard to multiagent systems: papers from the 1996 AAAI spring symposium 1996
1569
1513 guarantee that the specification of a system that has been pp.28–32. 1570
1514 designed actually fulfils the design requirements. Especially [13] Flores-Mendez R. Towards a standardization of multi-agent system 1571
1515 for critical applications, for example in real-time domains, frameworks, http://www.acm.org/crossroads/crew/roberto_flores-men- 1572
1516 there is a need to prove that the system being designed will dez.html. 1573
have certain properties under certain conditions (assumptions). [14] Geri S, Zhu Q. dbAgent: an intelligent web agent for database mining
1517 1574
International conference on computer and informatics (CS I’98) 1998
1518 Many popular multi-agent systems of today deploy agents in a pp.460–70.
1575
1519 uniform space of operating. The agents are supposed to respond [15] Goldman C, Rosenschein J. Emergent coordination through the use of 1576
1520 to the same calls and cooperate at the same time toward the cooperative state-changing rules Proceedings of the twelfth national 1577
1521 goals of operation. That kind of architecture is useful for some conference on artificial intelligence 1994 pp. 408–13. 1578
1522 applications. However, it endues some difficulties in agent [16] Giampapa J, Paoluc M, Sycara K. Agent interoperation across multi agent 1579
communications and task control. When applied in complex system boundaries Proceedings of agents 2000, Barcelona, Spain, June
1523 1580
3–7 2000.
1524 real-time situations with intensive human and system inter- [17] Grefenstette J, Daley R. Methods for competitive and cooperative co-
1581
1525 actions, the cooperative nature makes the system less robust evolution Adaptation, coevolution and learning in multiagent systems: 1582
1526 because the disability of one agent would affect the successive papers from the 1996 AAAI spring symposium 1996 pp. 45–50. 1583
1527 operations of the entire agent assembly. In this paper, we [18] Haddadi A. Towards a pragmatic theory of interactions Proceedings of the 1584
1528 studied four major architectural topologies of MAS. The first international conference on multi-agent systems (ICMAS-95) 1995 1585
1529 advantages and disadvantages of the topologies are assessed pp. 133–9. 1586
[19] Hayes-Roth B, Brownston L, van Gent R. Multiagent collaboration in
1530 and compared by using a set of criteria based on the 1587
directed improvisation Proceedings of the first international conference on
1531 functionalities and properties of agents in MAS. The study multi-agent systems (ICMAS-95) 1995 pp. 148–54. 1588
1532 and understand the MAS topology would help the effort of [20] Heinze C, Goss S, Josefsson T, Bennett K, Waugh S, Lloyd I, et al. 1589
1533 standardizing agent technology, and hopefully, promote more Interchanging agents and humans in military simulation. AI Mag 2002; 1590
1534 adoption of MAS in solving real world complex problems. 23(2):37–47. 1591
1535 [21] Hicks J, Stoyen A, Zhu Q. Intelligent agent-based software architecture 1592
for combat performance under overwhelming information inflow and
1536 1593
uncertainty Seventh IEEE international conference on engineering of
1537 7. Uncited references complex computer systems 2001 pp. 200–10. 1594
1538 [22] Lu H, Sterling L. SportsAgents: a mediator-based multi-agent system for 1595
1539 [10], [12], [17], [29]. cooperative information gathering from the world wide web Proceedings 1596
1597 of the fifth international conference on practical applications of intelligent Embodied conversational agents. Cambridge, MA: MIT Press; 2000. 1654
agents and agent methodology 2000 pp. 331–4. p. 95–122.
1598 1655
[23] Lu H, Sterling L. Intelligent matchmaking for information agents [29] Sahota MK. Reactive deliberation: An architecture for real-time
1599 cooperation on the world wide web Proceedings of the agent-based 1656
intelligent control in dynamic environments Proceedings of the twelfth
1600 simulation workshop 2000 pp. 161–8. national conference on artificial intelligence 1994 pp. 1303–8. 1657
1601 [24] Martin DL, Cheyer AJ, Moran DB. The open agent architecture: a [30] Shehory O, Kraus S. Task allocation via coalition formation among 1658
1602 framework for building distributed software systems. Appl Artif Intell autonomous agents Proceedings of the fourteenth international joint 1659
1603 1999;13(1–2):21–128. conference on artificial intelligence 1995 pp. 655–61.
1660
[25] Nwana HS, Ndumu DT, Lee LC, Collis JC. ZEUS: a toolkit for [31] Suzuki N, Fukuda M, Bic LF. Self-migrating threads for multi-agent
1604 building distributed multi-agent systems Proceedings of the third 1661
applications International workshop on cluster computing (IWCC’99)
1605 international conference on autonomous agents (Agents’99) 1999 1999. 1662
1606 pp. 360–1. [32] Sycara K, Klusch M, Widoff S, Jianguo L. Dynamic service matchmaking 1663
1607 [26] Perry B, Taylor M, Unruh A. Information aggregation and agent among agents in open information environment. ACM GISMOD Rec 1664
1608 interaction patterns in infosleuth Proceedings of CIA 99. New York: 1999;28(1):47–53.
1665
ACM press; 1999. [33] Virdhagriswaran S, Osisek D, O’Connor P. Standardizing agent
1609 1666
[27] Petrov PV, Zhu Q, Hicks JD, Stoyen, , AD. A hierarchical collective technology. ACM Stand View 1995;3(3):96–101.
1610 agents network for real-time sensor fusion and decision support The [34] White JE. Mobile agents. In: Bradshaw JM, editor. Software agents. 1667
1611 AAAI/KDD/UAI-2002 joint workshop on real-time decision support and Menlo Park, California: AAAI Press; 1997. p. 437–72. 1668
1612 diagnosis systems 2002 pp. 73–4. [35] Wilkins DE, Myers KL. A multiagent planning architecture Proceedings 1669
[28] Rickel J, Johnson WL. Task-oriented collaboration with embodied of the 1998 international conference on AI planning systems, Pittsburgh
1613 1670
agents in virtual world. In: Cassell J, Sullivan J, Prevost S, editors. 1998 pp. 154–62.
1614 1671
1615 1672
1616 1673
1617 1674
1618 1675
1619 1676
1620 1677
1621 1678
1622 1679
1623 1680
1624 1681
1625 1682
1626 1683
1627 1684
1628 1685
1629 1686
1630 1687
1631 1688
1632 1689
1633 1690
1634 1691
1635 1692
1636 1693
1637 1694
1638 1695
1639 1696
1640 1697
1641 1698
1642 1699
1643 1700
1644 1701
1645 1702
1646 1703
1647 1704
1648 1705
1649 1706
1650 1707
1651 1708
1652 1709
1653 1710