ACM SIGSOFT Software Engineering Notes Page 1 September 2005 Volume 30 Number 5

Some Agent Theory for the Semantic Web

Leona F. Fass
P.O. Box 2914, Carmel CA 93921
<lff4 [AT] cornell [DOT] edu>

Abstract as a multi-agent system we look at sub-systems constructed of

We take the position that for any goal achievable on the Semantic Web-dwelling agents to fulfill specific tasks, or to solve specified
Web, there will be a “best” system of Web-dwelling software problems, as “behavioral realizations”. With this very theoretical
agents to realize that goal, and that such a system may be perspective we propose that any goal achievable on the Semantic
discovered effectively. The process of determining the “best” Web will have a “best” multi-agent system behavioral realization.
agent system may be overseen by a distinguished Manager Agent. We take the position that this realization may be constructed or
But with realistic time and space constraints, and the dynamic adapted effectively from Semantic Web-dwelling software agents,
nature of the Semantic Web, finding an approximating system may perhaps by a distinguished Manager Agent. The system agents
be acceptable. The approximation then may be adapted iteratively, may invoke and compose Web services, as needed, to achieve
to approach the ideal. We show that very practical researchers behavioral goals.
have looked at software agents and Semantic Web problems in a Since this theoretical result may disregard real constraints on time
similar way, determining approximating sub-optimal systems and and space, we next discuss the necessity of accepting
subsequently adapting them. Their applied research confirms that approximating Semantic Web-dwelling multi-agent systems.
theory provides a good foundation for practice. These can adequately fulfill tasks or effect problem solutions that
Keywords: Software agents, multi-agent systems, Web services, are “good enough”, relative to time, space, and information they
the Semantic Web, theoretical foundations. may obtain. We describe adaptation of such approximating
systems, when tasks, goals, information, or available components
Introduction change, or when anomalies or defects are found.
The aim of Semantic Web designers and developers is to enhance
Despite the abstract approach we take when considering these
the existing Web, supplying structure and meaning that will
issues we find other, very practical, researchers have looked at
facilitate machine-to-machine interaction. With agent and multi-
multi-agent systems and the Semantic Web in a similar way. We
agent system research and development integrated into Web
describe their related work with multi-agent system managers,
development, it should be possible for software agents to utilize
approximating results, iterative approaches to the “optimal”, and
the Web environment, fulfilling complex tasks that human users
constraints applied so that features of the dynamic Semantic Web
have specified. Given a human-specified task (e.g., “Schedule my
can be represented to software agents effectively. Certainly, we
appointments for the East Coast trip”, or “Get me contact
believe that there are problems for which optimal results may
information for the Cornellians and Penn grads who live in this
never be found. But we also believe, from the practical work we
area”) capable agents could produce a satisfactory result. They
have seen, that when considering agents and the Semantic Web,
would access machine-readable information, discover fulfilling
theory provides a good foundation for practice.
Web services and automate their use, making logical decisions and
invoking and composing services as needed. Research Motivating Our Approach
In inceptive articles about the Semantic Web James Hendler [13] We were motivated to consider theoretical problems related to the
and Tim Berners-Lee, James Hendler and Ora Lassila [2] Semantic Web and Web-dwelling software agents from our
described some of its proposed features including languages, analysis of some very interesting research in several related areas.
communicating ontologies and inferencing mechanisms. They These included fairly recent work in: emergent semantics;
provided illustrative examples of “agent-based computing” and the automated composition of Web services; deduced interaction;
role that agents can play in Web applications, interacting with each collaborative learning; agent coalition formation; and an agent’s
other and with human users. There is typically an Artificial locally-closed view of its world. Each bears some relation to our
Intelligence emphasis on Semantic Web and agent research. automata theory view of behavioral modeling, and each has
Agents often are depicted as lifelike, as they may navigate the convinced us to extend our theoretical work into problems of the
Web and interact with other agents or with people. (The latter developing Semantic Web .
aspect also brings a Human-Computer Interaction perspective to In describing emergent semantics Steffen Staab [22] noted the
the field.) While they may communicate, learn and adapt, these possibility that semantic links to Web pages could be inferred from
agents are really just software. And so, as well, is the Semantic observations of actions of human users. Luc Steels [23] described
Web. Hence the development of the Semantic Web and Web- knowledge representation for agents, so that they could interact.
navigating agents must involve substantial Software Engineering. Equivalent external representations that agents needed to
In a project as massive as the Semantic Web there is room, and a communicate did not imply equivalent internal representations. In
need, for multiple and coordinated research emphases. our own inductive inference research (such as described in
We ourselves take an automata theory approach to agents and the connection with model-based validation [5]), we had shown that
Semantic Web, motivated by some intriguing software agent- and the structure of a language (or device) could be inferred by
Web-related research we identify as theoretically grounded. Thus observing examples of behavior. We’d also shown that
we look at tasks to-be-fulfilled or problems to-be-solved on the behaviorally equivalent results, with different underlying
Web as behaviors to-be-realized. Viewing the Semantic Web itself structures, could be found. Thus we were interested in the concept
ACM SIGSOFT Software Engineering Notes Page 2 September 2005 Volume 30 Number 5

of inferring semantics, and additional features enabling diverse have applied our own theoretical approach to forming coalitions of
agents to communicate on the Semantic Web . software agents [7]. Here again there are very practical
applications: Soh and Tsatsoulis [21] have described real time
We learned about aspects of knowledge representation for the
coalition formation, with results applied to multi-sensor target
Semantic Web from Sheila McIlraith [15], who discussed
tracking.
language and logic needed to effect automated composition of
Web services. The goal of such efforts has been to make Web Finally, a somewhat theoretical view of a real Web condition is the
services agent-oriented, rather than human-oriented, so that agent’s local closed world (LCW) view of the Semantic Web,
invocation of services can be automated. Then software agents described by Jeff Heflin and Hector Munoz-Avila [12]. The Web
could discover information, execute processes, and make is an enormous and dynamic domain, and a software agent really
selections of actions to take. The satisfying results would be cannot know when it has completed an adequate search for
service compositions and interoperability. This seemed to us a information. The addition of LCW extensions to a Semantic Web
very natural way to look at problem solutions: decomposing language could provide Web-navigating software agents with
problems into more-easily-solved sub-problems; finding sufficiently complete information to conclude they’d done
component sub-solutions; recomposing components into the “enough”. This work has interested us because we found, in our
solution that results. own work, that perfect results could be obtained in theoretical
problem domains only because such domains could be constrained.
That solution components may already be available on the Web
But in real problems without domain constraints, we determined
was shown to us by Richard Waldinger [25]. He discussed
that approximate results could be acceptable results, for they might
deductive techniques to find agents appropriate for solving sub-
be “good enough” [5-8].
problems, and to compose them into groups that effect problem
solutions. His method employed a theorem-prover that, given We recognized the relationships between software agent research
available agent capabilities as axioms, produced a theorem that and our theoretical work, and between Semantic Web research
“glued” appropriate agents together to achieve a behavioral goal. and our theoretical work, as we have described above. Noting
He showed that this process may also detect anomalies and similarities, we decided to apply our automata theory perspective
information inconsistencies on the Web. In the theoretically- to agents and the Semantic Web. We are not designing the
oriented research we have conducted, we’d first prove that a Semantic Web or constructing it ourselves, but we believe there
problem solution (i.e., specified behavioral realization) did exist are applications of our theoretically-oriented work to Semantic
and then, that it could be constructed effectively from Web development. This is particularly the case in connection with
representative components. Thus we could see similarities in mechanisms for service invocation, coordinated communities,
Waldinger’s work. When, in our work, we’d had what we believed discovery and selection of services and choreographies, interaction
to be a realization (or an approximation) and tested it against a protocols and architectures supporting agents and services within
behavioral domain we, too, might detect anomalies. In our case we the Semantic Web. We are looking at problems similar to those
used the information to correct defects and find better behavioral considered by agent developers and Web designers; they and we
representations. As “theoretical” as such research may seem, it has just use different terminologies. In our automata world, these
important real world applications. E.g., the deductive anomaly problems are all problems of determining the structure of a
detection research described by Waldinger in a symposium on behavioral realization that will achieve a specified behavioral goal.
logic-based program synthesis [25] was funded by NASA (the
The Automata World and the Agents World
National Aeronautics and Space Administration), DARPA (the
Defense Advanced Research Projects Agency) and ARDA (the In our view, the agents world and the automata world are very
Intelligence Community’s Advanced Research and Development much alike. Both worlds involve taking complex problems,
Activity). We have been encouraged to find such recognition of decomposing them into simpler sub-problems, determining or
the fact that practical domains can benefit from theory. constructing realizing sub-solutions, and recomposing the results
into the complex problems’ solutions. Both worlds involve taking
Research involving collaborative learning agents, and describing a complex behavioral goal (the problem) and finding a behavioral
formation of agent coalitions, showed us that software agents can realization achieving that goal (the solution).
act together in groups and adapt, approaching a behavioral ideal. Now, in the automata world, solvable or (finitely) realizable
Many different agent collaboration techniques described in [24] problems are the behaviors of (finite) systems or devices. Device
might achieve this. All are more or less goal-directed learning components are “states” connected into sub-systems/devices; these
methods, determining “best states” and actions to chose. Software correspond to solutions to sub-problems. Reaching a “final state”
agent coalitions defined by Leen-Kiat Soh and Costas Tsatsoulis corresponds to determining a solution to a problem. In our
[21], and those discussed throughout [20], are formed so that analogous view of the agents world, agents solving sub-problems
groups of agents can work together solving particular problems or correspond to “states”. Agent interactions are like state-to-state
fulfilling particular tasks. These agents might use case based transitions. And, when a group of software agents or multi-agent
reasoning and reinforcement to evaluate utility of working together system solves a problem or completes a process, it is as if that
and might “argue”, “negotiate” or even “coerce” each other into group has reached a “final state”. Thus synthesizing an automata-
joining the task-oriented group. An initial sub-optimal coalition theoretic system or device that realizes a specified behavior is
may be formed, and then refined to approach an optimal result. equivalent to composing a group of agents into a multi-agent
We, too, have iteratively constructed behavioral models that aim to system that produces a behavior, or fulfills a task. It is well known
achieve a goal in the “best possible” way, and have described in automata theory that there is a relationship between the
feasibility of adaptations to software design [5]. Additionally, we components of a realizable behavior and the structure of its
ACM SIGSOFT Software Engineering Notes Page 3 September 2005 Volume 30 Number 5

realization. Behavioral elements can be grouped (based on tests, we know the PTA works “well enough”. I.e., within time
experiments that determine their “indistinguishable behavior”) into limits, with sufficient protein-rich energy bars, and without
classes that correspond to the components or “states” of a realizing exhausting our travel budget, we have arrived safely at the
“device”. Relationships among the behavioral classes define respective destinations. The specified behavioral goals have been
“state-to-state transitions”. Classes of “correct” behavioral achieved.
elements define final, or goal, states. In the automata world, just
And What About the Semantic Web?
by analyzing the components of a specified realizable behavior or
goal, one may determine a device that produces or achieves it. One If we consider the collection of software agents that may dwell on
may optimize a result by minimizing it and, for finitely realizable the (Semantic)Web, the Web itself is a multi-agent system. At least
behaviors, one may find the optimal result effectively. for our current purposes, that is how we view it. So, just like
Furthermore, if a finite potential realization is given, and the Waldinger’s approach [25] of locating agents on the Web and
domain of all “correct” and “incorrect” behavior is known, one deductively “gluing” them into problem-solving systems, our
may conduct effective tests to detect anomalies (and iteratively automata theory approach, too, can apply to the whole Semantic
correct them). Alternately, after testing and perhaps correcting, Web. When synthesizing a software agent “device” to fulfill some
one may determine the realization to be optimal. The components behavioral specification, agents will be collected from the
of an optimal result are found quite easily, just by exploiting the Semantic Web agent universe. The very same possibilities we
relationship between behavior and structure. described above (merger, minimization, elimination, optimization,
adaptation) that might apply to “one’s own personal multi-agent
Because of the similarities between the agents world and the system”, need only be scaled up to apply to the whole Semantic
automata world, we believe that well-known automata-based Web. In actuality this is, of course, quite a scaling project. But it is
techniques can be adapted to software agents. Then suitable agents all quite simple in theory.
may be collected into groups that can solve specified problems
effectively. By extending proven automata theory techniques to We believe that the best way to approach this project is to develop
processes in the agents world, agents may be collected into a distinguished Manager Agent. This agent would oversee the
interacting systems or “devices” that realize behavioral goals or selection, composition, and adaptation, etc., of the multi-agent
fulfill specified tasks. Once again, this just requires determining systems constructed from the other software agents dwelling on
components and architecture of an agent system from analysis of the Semantic Web. The Manager Agent would assign agents to
its proposed behavior. groups, based on classes of tasks and the agents’ abilities. The
Manager would experiment, to effect mergers and reductions. The
We believe that automata theory methods can be applied to Manager would learn of new agents and services and utilize them
systems of agents so that optimal (e.g. minimized) agent systems as warranted. Adaptation would occur as new tasks developed, too.
will be discovered to fulfill specified tasks. Utilizing this theory: In our inference work we had a theoretical informant (many call
useless agents in a system may be discovered and removed; agent this an “oracle”) who oversaw the construction and adaptation of
configurations behaving indistinguishably can be merged; and behavior-realizing models. This seems the ideal theoretical
communication paths can be minimized. We believe that when a solution for determining architecture of each goal-directed agent
behavioral goal grows or changes, automata theory techniques will system that could be composed, or exist, on the Semantic Web.
help collected agents (or, a multi-agent system) expand The Manager (acting as informant) would oversee, track, select,
monotonically or adapt as needed, with amended architecture construct and watch the Semantic Web grow and change.
and/or different “final states”. Additionally, successfully
performing agent sub-systems can be reused within future Of course there are time and space considerations relating to how
configurations. On the other hand, a multi-agent system or agent this Manager would work. There would be obvious intractability in
group can be tested to detect anomalies, if a goal behavior and its making “perfect” system plans. Real time solutions to most
complement can be finitely categorized. Once again, we propose problems could not possibly be obtained. Not if the Manager must
this can all be done, just by exploiting the relationship between the categorize classes of all Semantic Web-dwelling agent capabilities
goal behavior and the agent group, or multi-agent system, and collect them into systems fulfilling behavioral goals. Not if the
structure. Manager must find optimal structures formed from among all
other agents. Even with a Local Closed World assumption,
We have illustrated some of these claims with a Personal Travel keeping a Manager in touch with “much” information as the
Assistant (PTA): a collection of task-fulfilling software agents Semantic Web grows, the job could not, in reality, be done. But in
initially described as a theoretical construction in [7]. PTA Good Old Fashioned Automata Theory we may establish the
components that may plan our possible trip to a symposium in following theoretical result:
Washington DC could be adapted quite easily from the
For any finitely realizable behavioral goal achievable
components that planned our trip to the ICSE meeting in
on the Semantic Web, there will be a “best” (perhaps
Edinburgh, Scotland, or to an Information Sciences conference in
minimal, or most efficient, or other designated
Salt Lake City. All results could be optimized with respect to our
optimal) finite system of Web-dwelling software
established criteria (e.g., our specific constraints on finances, diet
agents to realize that goal. This system may be
and time). Components that plan our trips to meetings in San Jose
discovered effectively: by construction from existing
would be useless, and might be removed to save PTA space. (We
agents; or by adaptation from an agent system that is
travel to San Jose by bus and train; the other specified destinations
already known.
require taxi, airport van and plane.) We may not be able to finitely
characterize all possible “correct” or “incorrect” travel behavior The key words in the above assertion are “achievable” and
that the PTA might produce. But from our personal-experience “effectively”. If the goal is finitely realizable there will be a finite
ACM SIGSOFT Software Engineering Notes Page 4
“device” to achieve it, and the device will have finite components.
If the goal is achievable on the Web, the realizing-device
components will be Web-dwelling. Although the Web is dynamic
and growing, at every instant it is finite. So discovering these
existing component agents might require an enormous search, but
would be accomplished effectively. Determining the interactions
among the finite components to “construct” the “device” would
then be an effective process, too. Once a realizing system is
constructed from Web-dwelling agents, or an existing realizing
system is found, it may be optimized. (For example, redundant
components may be merged [7,13].) Comparisons among potential
realizing systems could select a “best” configuration based on
specified criteria, given the information that is available in that
instant. The Manager Agent would select the instant, and would
have access to information available at that time. (There is no
confrontation with the classical automata theory “Halting
Problem”.) When the Manager has concluded sufficient
construction work has been done, the result would be accepted as
“good enough”.
The construction and selection processes may seem tedious and
may be impossible to complete in real time, but they could be
completed effectively. And with a distinguished Manager Agent to
do the work this is a nice result, in theory. Furthermore, it provides
a basis for determining approximating goal-directed Semantic
Web multi-agent systems that are quite acceptable, in practice.
Knowing that that an optimal result exists, but that its discovery
may not be achieved in a reasonable amount of time (or space),
makes one more willing to accept an approximation to the ideal
that can be determined feasibly.
Related Work
Above we provided the theory that if something can be achieved
effectively on the Semantic Web, then a multi-agent system may
be constructed effectively from Web-dwelling components to
achieve it. A Manager Agent can complete the construction work,
assuming adequate time and space. If a “perfect” result exists,
then we know that, with sufficient time and space, that result can
be obtained. But we accept an approximation to a behavioral ideal
as a “good-enough” result when time and space are constrained.
Finding an approximating multi-agent system that achieves “good
enough” results, or that achieves “perfect” results “enough” of the
time, in reality is satisfactory. Should construction defects or
behavioral anomalies be detected when the approximating multi-
agent system performs, the system may be revised. The Manager
can observe the performance behavior (perhaps even with input
commentary from users), and adapt the multi-agent system
structure. With each iteration the system will approach, more
closely, the behavioral ideal. This iterative testing-and-adapting is
precisely what is done, or should be done, when a complex
software system is deployed [5]. Inevitable defects can be detected
and a better, “more correct” complex software system may result.
One could hardly find a more complex software system than the
Semantic Web.
Many practical researchers have found approximating multi-agent
systems as satisfactory results, and have developed iterative
processes for their subsequent adaptation. Some have constructed
systems from categories of locally available agents, so that locally
adequate (but globally approximate) results are found. Some have
employed supervisory agents to oversee the activities of other
September 2005 Volume 30 Number 5
agents, teaching or guiding them to group together and perform.
Adaptation of initial configurations, or of supervisor behavior,
often has been described. We, and these diverse researchers, all
seek the best possible results with respect to multi-agent systems
and a dynamic Semantic Web, even if using differing techniques
and terminologies.
A managerial “teacher agent” is utilized by Sandip Sen and Parijat
Prosun Kar [19] sharing knowledge with other (“student”) agents
and employing varying information-conveying techniques:
instance-based knowledge; decision trees, etc. Markov Decision
Processes (MDPs) are employed by Patrick Riley and Manuela
Veloso [17], to be utilized by a “coaching agent” that manages
other software agents. An MDP is learned based on the coach’s
observations of past performance. Then an optimal policy may be
determined, and used by the coach to advise the other agents on
future actions to take. This work was first developed in a simulated
soccer-game domain, and then extended to more general domains.
The authors have found there is an advantage to separating the
“coach” from the “team” of agents that the coach may advise.
E.g., the designated coach can work with different teams and
develop different strategies, as needed, for each. This would
certainly be advantageous for our proposed Manager Agent,
overseeing the Semantic Web. An entire multi-agent system is
viewed as an MDP by Carlos Guestrin, Shobha Venkataraman and
Daphne Koller in [11]. Value functions (approximations to the
optimal) are computed for an agent, based on current state and
actions of some subset of the remaining system agents. The agents
communicate via a coordination graph, to decide on optimal (with
respect to the value functions) joint actions to take. The process is
reiterated, potentially producing a different coordination action at
each step of the plan. This iterative approach has been applied to
multi-agent system network administration problems, and the
authors have found it to be substantially more efficient than other,
e.g. tabular, methods. We would interpret the repository of the
successive coordination graphs to be the multi-agent system
manager.
A Bayesian modeler software agent is described by Leonardo
Garrido, Ramon Brena and Katia Sycara [9], to build probabilistic
models of its “competitor agents” in an iterative way. The
modeler’s decision-theoretic approach chooses a rational decision
at each round of a “game”, maximizing modeling ability with
respect to gain of its “most dangerous opponents”. A Bayesian
updating mechanism builds models about the others after each
round, iteratively and incrementally. The authors note that with
iteration, the modeler’s performance becomes close to that of an
oracle. They conjecture that these Bayesian modeling methods are
optimal, with respect to information that is available. Here, too, a
manager agent is used; iterations of approximating solutions are
used; and an optimal result is obtained, when realistic constraints
are applied to a dynamically-changing domain. Carlos Guestrin,
Michail Lagoudakis and Ronald Parr’s software agent systems
[10] employ coordinated reinforcement learning techniques.
Agents find jointly optimal actions without having to consider all
possible actions in their joint, exponentially large, action space.
Action selection activities and parameters are iteratively updated,
so that both the “learning” and “teaching” are, relatively, optimal.
We noted earlier that reinforcement and case-based reasoning were
described by Leen-Kiat Soh and Costas Tsatsoulis [21], and used
when initial sub-optimal, goal-directed agent coalitions were
ACM SIGSOFT Software Engineering Notes Page 5 September 2005 Volume 30 Number 5

adapted iteratively. Xin Li and Leen-Kiat Soh [14] employ Still, despite the successful research we have seen, we realize there
reinforcement for agent coalition formation, as well. Again, are agent and Semantic Web development issues that can never be
optimal behavior is approached in a dynamic environment, based resolved [6,8]. Among such issues are those relating to
on the information that is available at each particular time. information verification and trust. There is no way to control the
accuracy of the information that may be placed on the Semantic
Agents in the congregations described by Christopher Brooks and
Web, so a navigating software agent is dependent on what is there.
Edmund Durfee in [1] have “self-organized”, and they have done
And, in addition to issues of tracking the dynamic Web, there is
so to reduce a plethora of information. By forming congregations
also an issue of dynamic problems. I.e., in many real life problems,
within a multi-agent system, member agents can find other
once a problem solution is determined, the problem parameters
congregation-member agents that are suitable to work with on
may change (The Fass-Littman Theorem of Life: “Things always
specific tasks. This enables these agents to ignore the global state,
happen”). Perhaps, in such cases, optimal results may never be
directing their attention to the local state of the congregation.
found.
Suitable partners are then found at low computational cost. The
Brooks and Durfee congregations are closely related to our own However, the many examples of applied research that we have
automata-theoretic notion of behavioral classes corresponding to described above use methods similar to those we have employed in
components in a realizing system or device. However, the agent our own theoretical approach, and they have been successful. We
congregations that are described are formed by agent preferences. conclude that the work cited demonstrates: when considering
The authors admit that what may be gained in cost-effectiveness, software agents and the Semantic Web, theory provides a good
may be lost in utility. foundation for practice.
Finally, Organization Theory is used by Scott DeLoach and Eric
Matson [3] to design adaptive multi-agent systems. The elements Acknowledgments
of their model include goals, roles, and agent capabilities. Their We are grateful to Sheila McIraith, Jim Hendler, Richard
adaptive information system can select the best agents to assign to Waldinger and Jeff Heflin, whose understandable explanations
appropriate roles so that specified goals may be achieved. From interested us in the Semantic Web and showed us it is all quite
our perspective, this is just like determining structural components logical. Discussions with Michael Littman about real applications
of a device that realizes a specified behavioral goal. To us, the of theoretical results led to “The Fass-Littman Theorem of Life”.
information system is the theoretical informant, or multi-agent We were invited to participate in the AAAI 2005 Fall Symposium
system manager, that assists in construction of the realizing on Agents and the Semantic Web, based on the initial draft of this
“device”. DeLoach and Matson have demonstrated their approach paper. Remarks by anonymous symposium referees, and
with applications to a military domain. In an example they provide subsequent suggestions by Will Tracz, assisted us in completion of
a system might be asked, say, to provide a list of enemies moving this improved and expanded version.
vehicles within a specified geographic area. To respond, the
system would need to activate appropriate software to apply
sensors, access data bases, and so forth: considering goals, roles References
and capabilities. [1] Brooks, C.H. and E.H. Durfee (2002): “Congregating and
Once again, we see that theory can be applied in very practical Market Formation” in [24], pp. 33-40.
domains. If we view the Semantic Web as a multi-agent system [2] Berners-Lee, T., J. Hendler and O. Lassila:(2001): "The
containing “all possible agents”, theory may be applied Semantic Web", Scientific American , May 2001, pp. 34-43.
“everywhere”.
[3] DeLoach, S.A. and E. Matson (2004): “An Organizational
Conclusions Model for Designing Adaptive Multiagent Systems”, in [4], pp.
We have described motivating research involving software agents 66-73.
and the Semantic Web that led us to look at Semantic Web [4] Dignum, V., D. Corkill, C. Jonker and F. Dignum, Eds.
development from the perspective of automata theory. Taking the (2004): Agent Organizations: Theory and Practice, Papers from
Semantic Web as multi-agent system itself, we have provided the the AAAI Workshop, San Jose CA, July 2004, AAAI Press WS04-
theoretical result that, from therein, a sub-system of agents may be 02.
constructed effectively to realize any specified (Web-achievable)
behavioral goal. Furthermore, at any instant a “best” such system [5] Fass, L.F.(2001): “Determining Software Models that are Less
may be found. We have suggested that system construction may be Incorrect”, Model Based Validation of Intelligence, Papers from
effected by a distinguished Manager Agent, observing and the AAAI Spring Symposium, Stanford University, March 2001,
overseeing all the other available agents. We have also noted that AAAI Press, SS01-04, pp. 113-116.
due to time and space constraints and the dynamic nature of the [6] Fass, L.F.(2002): “Can We Improve Web Access in the Real
Semantic Web, constructing an approximating system, and then World?”, appears as Statement of Interest at URL:http://
adapting it, would be a satisfactory result. To show that theory can
projects.teknowledge.com/AAAI-2002/Fass.html and in [16], pp.
be applied, we have referenced our Personal Travel Assistant [7] xv-xvi.
and, more so, the work of numerous practical researchers. They
have employed varying techniques and models to find [7] Fass, L.F. (2004): “An Automata-Theoretic View of Agent
approximating systems, and have adapted them to approach those Coalitions”, in [20], pp. 18-21.
that are optimal.
ACM SIGSOFT Software Engineering Notes Page 6 September 2005 Volume 30 Number 5

[8] Fass, L.F. (2005): “Ubiquity and Common Sense for the
Web”, in Proceedings of the Joint Conference on Information
Sciences/Workshop on Web Intelligence and Security, Salt Lake
City UT, July 2005, pp. 1433-1436.
[9] Garrido, L., R. Brena and K. Sycara (2002): “On Measuring
the Usefulness of Modeling in a Competitive and Cooperative
Environment” in [24], pp. 48-54.
[10] Guestrin, C., M. Lagoudakis and R. Parr (2002):
“Coordinated Reinforcement Learning”, in [24]: pp. 98-105.
[11] Guestrin, C., S. Venkataraman and D. Koller (2002):
“Context Specific Multiagent Coordination and Planning with
Factored MDPs”, in [24], pp. 17-24.
[12] Heflin, J. and H. Munoz-Avila (2002): “LCW-Based Agent
Planning for the Semantic Web”, in [16], pp. 63-70.
[13] Hendler, J. (2001): “Agents and the Semantic Web”, IEEE
Intelligent Systems, March/April 2001, pp. 30-37.
[14] Li, X. and L-K Soh (2004): “Investigating Reinforcement
Learning in Multiagent Coalition Formation”, in [20], pp. 22-28
[15] McIlraith, S. A. (2002): “Automated Composition of Web
Services”, Invited Symposium Talk, AAAI Spring Symposium on
Logic-Based Program Synthesis: State of the Art and Future
Trends, Stanford University, March 2002
[16] Pease, A., R. Fikes and J. Hendler, Eds. (2002): Ontologies
and the Semantic Web, Papers from the AAAI Workshop,
Edmonton AB, July 2002, AAAI Press, WS02-11.
[17] Riley, P. and M. Veloso (2004):“An Agent Coaching
Agents”, in [18], pp. 61-62
[18] Rosenstein, M. and M. Ghavamzadeh, Eds. (2004):
Supervisory Control of Learning and Adaptive Systems, Papers
from the AAAI Workshop, San Jose CA, July 2004, AAAI Press
WS04-10.
[19] Sen, S. and P.P. Kar (2002): “Sharing a Concept”, in [24],
pp. 55-60.
[20] Soh, L-K and J.E. Anderson, Eds.: (2004): Forming and
Maintaining Coalitions & Teams in Adaptive Multiagent Systems,
Papers from the AAAI Workshop, San Jose CA, July 2004, AAAI
Press, WS04-06.
[21] Soh, L-K and C. Tsatsoulis (2002): “Learning to Form Nego-
tiation Coalitions in a Multiagent System”, in [24], pp. 106-112.
[22] Staab, S. (Ed) (2002):“Emergent Semantics”, IEEE
Intelligent Systems, January/February 2002, pp. 78-86.
[23] Steels, L. (2002):“Language Games for Emergent Semantics”
in [22], pp. 83-85
[24] Tumer, K. and P. Stone, Eds. (2002): Collaborative Learning
Agents, Papers from the AAAI Spring Symposium, Stanford
University, March 2002, AAAI Press, SS-02-02.
[25] Waldinger, R.(2002): “Deductive Chat Lines for Multiple
Agents”, Invited Symposium Talk, abstracted in Logic-Based
Program Synthesis: State of the Art and Future Trends, Papers
from the AAAI Spring Symposium, Stanford University, March
2002, AAAI Press, SS-02-05, p. 5.