Modeling Agents and their Environment:
The Communication Environment
James Odell,
James Odell Associates, Ann Arbor, U.S.A.
H. Van Dyke Parunak and Mitchell Fleischer,
Altarum, Ann Arbor, U.S.A.
|
 |
COLUMN
PDF Version |
Abstract
Without an environment, an agent is effectively useless. Cut off from
the rest of its world, the agent can neither sense nor act. An environment
provides the conditions under which an entity (agent or object) can
exist. It defines the properties of the world in which an agent will
function. Designing effective agents requires careful consideration
of both the physical and communicational aspects of their environment.
This second part of this paper and deals with the communication environment.
1 INTRODUCTION
Agents need to operate and exist within an environment. Figure 1 illustrates
a common view that agents perceive their environment though sensors
as well as effect actions on it. [Pfeifer, 1999; Weiss, 1999; Russell,
1995] For example, a Stock agent can receive an event indicating that
quantities of a particular part are low. The agent then decides whether
more parts need to be ordered and, if so, put out a general call-for-proposal
so that interested vendors can reply. When proposals arrive, the Stock
agent will choose and notify the winning vendor. This model implies
that agents interact via an environment. Even direct communications
(such as vender notification) must occur through some medium. In other
words, the environment provides the appropriate conditions that enable
interaction among agents. This insight, largely overlooked in the design
of purely electronic agents, is particularly critical for managing agents
that are situated in the physical world.
Fig. 1. Agents interact with and through their environment.
2 WHAT IS AN ENVIRONMENT?
One of the key properties of agents is their autonomy. However, autonomy
is not an all-or-nothing issue. Practically speaking, agents can neither
be totally free of external dependencies nor completely reliant on them.
They always depend on external factors to some degree.
| An environment provides the conditions under
which an entity (agent or object) exists. |
In other words, it defines the properties of the world in which an
agent can and does function. An agent’s environment, then, consists
not only of all the other entities in its environment, but also those
principles and processes under which the agents exist and communicate.
Designing effective agents requires careful consideration of all of
these factors when designing their environment.
A canonical example of agents situated in an environment is an ant
colony. Ants interact with one another largely through chemicals, called
pheromones, that they deposit in the environment and then sense
to guide their actions. Numerous individual interactions yield the emergent
development of paths through the environment. However, the environment
is more than just a communication channel. Agents depend both on tangible,
physical support and on other agents. Two aspects, then, are critical
for agent environments (and the formation of paths): the physical and
the communicational. The previous paper discussed with the physical
environment; this paper discusses with the communication environment.
3 COMMUNICATION ENVIRONMENT
In individualist agent environments, agents are viewed as
independent entities; whereas in collectivist environments,
agents are viewed as interdependent. While an agent can operate by alone,
the increasing interconnections and networking require a different kind
of agent—one that can communicate effectively with other agents.
A communication environment provides two things. First, it provides
the principles and processes that govern and support the exchange of
ideas, knowledge, information, and data. Second, it provides those functions
and structures that are commonly employed to enhance communication,
such as roles, groups, and the interaction protocols between roles and
groups. In short:
| The communication environment provides those
principles, processes, and structures that enable an infrastructure
for agents to convey information. |
Communication, Interaction, and the Social Agent
Communication
Basically, communication is the conveyance of information from one
entity to another. The nature of this transfer can range from the simple
to the complex. For example, a satellite could periodically send one
bit to inform ground control that it is still functioning correctly;
in contrast, the information exchanged within the US Senate to negotiate
tax cuts can appear quite chaotic. In contrast, broadcasts such as television
commercials do not necessarily result in communication. A signal may
go out, but if you are not listening or watching, how can the commercial
convey information?
(a) Two agents with no communication activity.
(b) One agent transmitting to another, but not communicating.
(c) One agent communicating with another agent (but
not interacting).
Fig. 2. Agent transmission versus communication.
Figure 4 illustrates the difference between transmission and communication.
In Fig. 2(a), neither agent has any transmission activity. Figure 2(b)
indicates that the agent on the left transmitted information through
the environment, but was not received by the other agent. Communication,
however, requires that the information transmitted by one agent results
in a state change of another (Fig. 2(c)). In the case of television
commercials, perceiving its transmission means that your senses have
at least detected it. The perception could involve you buying the advertised
goods, throwing a shoe at the television screen, or simply choosing
to do nothing. Either way, communication has occurred because the act
of sensing and deciding involves a state change by the receiver.
Interaction
(a) One agent communicating with another agent; and
the other agent transmitting (b) a response, but not communicating or
interacting.
(b) Two agents interacting interacting.
Fig. 3. Agent communication versus interaction.
Proving that communication has occurred, however, requires
us to know that the inner state of the receiving agent has in fact changed.
We are not advocating that the communication environment possess such
mentalistic knowledge—only that such an environment be present
so that transmission and communication can occur. However, knowing that
a transmission was received can be important to the sending agent. One
useful way to determine if communication has occurred is when an interaction
results. Figure 3(a) depicts one agent communicating with another. Here,
the other agent responds, but the original agent does not receive the
responding transmission. (The original agent, then, cannot know for
a certainty whether communication occurred.) In other words, there was
no interaction between the two agents. Interaction1
requires two-way communication (i.e., a reciprocal effect), as illustrated
in Fig. 3(b). Interaction, then, not only defines exchange of information,
it confirms that the original transmission was in fact received by the
other agent. In other words, the original agent can infer that its transmission
was communicated to the other agent as soon as a response is received—even
if the response communicates only that the responder did not understand
the original message.
Social Environment
In agent-based systems, communication and interaction are commonly
employed together. Furthermore, agent-based communication can even involve
patterns of interaction, or interaction protocols. From simply
requesting the price of a product to conducting elaborate contract-bidding
activities require that some agree-upon approach be in place to facilitate
interactive communication—without which the conveyance of information
could easily result useless Babel. Such a situation could be considered
social.
| A social environment is a communication
environment in which agents interact in a coordinated manner. |
As illustrated in Fig. 4, the social environment is a subset the communication
environment. In other words, not all communication is social (as defined
above), but all social activity requires communication.
Principles of a communication environment
Communication principles provide us with the fundamentals that are
essential for interactions, customs, norms, values, commitments, dependencies,
and so on, that constitute an agent society. The canonical ant’s
communication environment is simple: all communications between ants
are via pheromones. Here, the communication involves a two-step process:
the ant deposits pheromones that act as information for other ants,
while the “other ants” query the local environment for the
presence of pheromones. In short, these ants participate in a social
communication by way of environmental substances, rather than directly
with each other. Additionally, some ant societies have multiple kinds
of pheromones: one for exploring and one when returning with returning
food. The net effect is informing a society of ants about how to find
food or home—whichever is useful for any given ant. If an ant
is foraging, information about where to find food is useful; if the
ant has food, directions on how to get back the colony would be useful.
Such interaction is social because it provides the ants with an infrastructure
for the colony because it produces coordination among the ants.
Supply-chain agents can have elaborate collaborative protocols acquiring
and delivering goods and services along value-adding chains. Defense-related
protocols, require different interaction policies at different command
levels. Both direct and indirect interaction can be employed as interaction
strategies in thee applications.
Fig. 4. Social environment: coordination—and possibly
cooperation, and competition
In rich multiagent societies (MAS), several principles are required
to facilitate the communication environment.
- Communication language - Agents communicate to
understand and be understood. The formal study of communication has
three aspects: syntax, semantics, and pragmatics. Agent-based social
environments must define the principles required to address these
aspects. Additionally, it must define the types of messages that will
be employed (e.g., assertions, queries, replies, requests, and denials)
and the ontology. Some of the common agent communication languages
(ACL) languages include FIPA ACL, and KQML.
- Interaction protocols - An agent interaction protocol
(AIP) describes a communication pattern as an allowed sequence of
messages between entities and the constraints on the content of those
messages [Odell, 2000a and 2000b]. Examples of AIPs include the contract
net protocol, Dutch auction protocol, and publish/subscribe protocol.
FIPA has standardized more than a dozen AIPs [FIPA, 2000].
- Coordination strategies - Agents communicate to
achieve their goals and the goals of the social group in which they
participate. Cooperation, competition, planning, and negotiation are
common principles used to perform activities in a shared environment.
AIPs can be associated with each of these strategies.
- Social policies - The permissions and obligations
that dictate acceptable social behavior. They include being able to
apply and enforce these policies across distributed agents and systems.
The general focus here is on the application and management of policies
on agents and groups of agents—not the detailed management of
agent lifecycles and areas currently addressed by FIPA agent management
specifications. Other considerations for social policy can involve:
- Implicit vs explicit rules; not all rules are specified
in advance: i.e., learning what the rules are or adjusting to
a change in rules, emergence of rules, unconscious rules, when
tradition becomes a social norm, or policy.
- Different levels: of influence/power (e.g., the ROI on obeying
or violating a rule, strength and “evaporation” of
rule; rules don’t always stick around, rules as memes, language
use.)
- Culture - a set of values, beliefs, desires, intentions,
trust, morality. These can determine the characteristics of the above.
FIPA vs. KQML cultural differences; English vs. other different-culture
language (e.g., Navajo). Culture also affects language, interaction
protocol, and social policies (implicit & explicit).
Processes of a communication environment
An agent’s communication environment provides processes that
enable agents to interact productively2. In particular, it
must provide:
- Interaction management - managing the interactions
among entities to ensure that they are adhering to the selected agent
interaction protocol (AIP). AIP adherence can be maintained by those
agents participating in the protocol, so that the environment does
not need to be involved. However, trusting that each agent can and
will adhere to and ensure correct AIP interaction may not be enough
to ensure social order. An environment-level control can be implemented
as an AIP-manager agent. Did you get what you wanted/needed/expected;
- Language processing and policing - where the language
parses correctly, it parses correctly but is wrong (evidence or contradictory),
or is correct but inappropriate within the agent’s context.
- Coordination strategy services
- Directory service - locating agents can be
supported by white-page (individual), yellow-page (industry),
or green-page (offered services) methods. In the physical environment,
this directory is used to provide information about where the
agent is physically; in the social environment, it provides information
about an agent’s role or the services that it can provide.
- Mediation services - acting through an intermediate
agency. Specialized agents could be established in the environment
to act as a communication’s intermediary for activities
such as transaction management or ontology translation. Environment-level
mediation can be implemented using specialized agents.
- Policy enforcement service - control of the agent
by its environment or social group. The range of possible mechanisms
for enforcing policy mechanisms can range from social sanctions to
a complete withdrawal of supporting services for the non-conforming
agent.
- Social differentiation - the process whereby a
group or community becomes separate or distinct. To ensure success,
groups will institutionalize and employ roles for their members. An
agent can play multiple roles in multiple groups.
- Social order - the production of a structure of
relationships among social agents [Castelfranchi, 2000]. Social order
can be the result of formal policies as well as emerge via self-organizing
mechanisms. The later is a emergent social pattern of its own, such
as the stock market. The former has to do with managing the conditions
of an agent society as a whole employing a non-accidental and non-chaotic
pattern of interactions. For example, auctions employ strict social
patterns. Such a mechanism can be employed to control undesirable
emergent patterns that need to be remedied. For example, when stock
prices rise or fall by too many points in a session, trading curbs
are triggered.
Population of a communication environment
As mentioned earlier, a physical environment consists of all those
entities in the physical environment. In contrast, an agent’s
social environment consists of
- those social units in which the agent participates,
- the roles that are employed for social interaction,
- all the other members who play roles in these social units.
Each social unit, or group, is a set of agents associated
together by some common interest or purpose. There are three reasons
for creating groups.
- Intragroup associations - Groups are commonly
formed to foster or support the interaction of those agents within
the group. Here, the group provides a place for a limited number of
agents to interact among themselves. For example, such agents might
wish to exchange information or seek safety in numbers.
- Group synergy - Social units can be formed to
take advantage of the synergies of its members, resulting in an entity
that enables products and processes that are not possible from any
single individual. Corporations, unions, and governments are examples
of such social units.
- Intergroup associations - Social units also serve
as an entity with interactive capability. Here, a group is a set of
agents that interact with other sets of agents. Recurrent patterns
of interaction define roles, and frequently associated roles are usually
considered as defining (sub)groups.
A group can be empty if no agents participate in the group; its collection
can also contain a single participating agent or multiple agents. Groups
have a separate identity within a larger whole and can be composed of
agents, as well as other groups3. Furthermore, groups can
become social actors influencing group processes and outcomes, as well.
For example, most business organizations interact with sector groups
such as industry, technology, agriculture, and government; and each
of these can influence the other as well as consist of their own subgroups.
In this way, an agent social environment can be thought of as a society
where agents interact in a more or less ordered community.
A role is an abstract representation of an agent’s
function, service, or identification within a group. In other words,
each role is a class of agents that participates in pattern of dependencies
and interactions in a prescribed manner. A pattern of dependencies is
an important component of a role. For example, if agent A is a customer,
there must be some agent B on whom A depends for goods and services,
while B depends on A for money. For AIPs, roles define which actions
are permitted for a certain class of agents. For example, an agent playing
the customer role may request goods, but not supply them; the supplier
has the opposite requirements. [Parunak, 2001]
4 SPATIAL AND TEMPORAL CONSIDERATIONS
An agent’s environment – physical or social – must
occupy both space and time. Agent populations abide and interact, their
processes occur, and their environmental principles are defined over
that same temporal space. Agent space and time involves the notion of
agent place, along with two of its primary attributes: extent,
and locality.
Place
Each agent environment can be thought of as a whole or it can be subdivided
into discrete regions. Regions partition the agent’s
physical environment into smaller physical units – where each
region may have different or unique characteristics. For example, a
grid structure can be defined for the ants so that discrete locations
are provided for both the ants and their pheromones, as well as the
ability to form pheromone paths. Region definition can also include
geographic-based attributes, such as lakes, hills, roads, and structures.
In social environments, regions spatially partition the environment
into groups and roles. In contrast, temporal space can define unique
characteristics for each place in time.
Region specification can include various constraints. For example, in
a physical environment we may wish to specify that no two ants may occupy
the same place at the same time; yet, we may permit accumulation of
multiple units of pheromones. In a social environment, business organizations
might be limited to having one person occupying the role of president
at any point in time.
Set theoretic distinctions can be made between membership and set.
For example, if Agent A belongs to Organization B and Organization B
belongs to a federation of organizations C, A does not necessarily belong
to C. However, if an Agent A is an element of Set B and Set B is contained
by Set C, then Agent A is also an element of Set C.
The region size is determined based on the design granularity: meter-sized
places are unrealistic for small ants; micron-sized regions would push
the limits of current technology. For example, Pacific Gas & Electric
specifies a longitude and latitude within two meters accuracy called
a geocode. The geocode place size for PG&E, then, is four-square
meters. In combat examples, a similar grid structure and size is also
employed.
In another example, SRI proposes a new top-level Internet domain called
.geo. [SRI, 2000] In a .geo system, the Earth would be partitioned into
cells based on latitude and longitude. Dedicated servers would hold
the data registered to Web sites within its geographical domain, as
well as maps and other information. As illustrated in Fig. 5, places
can be arranged hierarchically so that search engines could direct queries
to one type of server, depending on what the Web user was looking for.
The Internet user could then query for cardiac bench surgery in North
America or men’s clothing stores in Ann Arbor, Michigan. In this
way, web user would never need to use unwieldy .com addresses; the geo-enabled
search engine translate a geographic location into web sites registered
at that location.
Fig. 5. Places has a hierarchy of geographic placement. [SRI, 2000]
Extent
Agent environments must exist in some designated area (or volume)
in space and time. Region designations can be expressed in various ways:
length/width/height, location points indicating the boundary, memory
or disk locations, to and from dates, and so on. The shape of physical
space can also be considered here. Social space can expressed in terms
the degree of interaction. For example, this could include the number
of people you work with, the “degree of separation” between
one website and any other website. For example, the environment could
be a flat plane or a torus space. In other words, agent environments
require an extent that defines its size, shape, and boundaries. Effects
of boundary conditions can also be addressed here.
Locality
The ability to locate an entity is an important factor, particularly
in agent communication. Locality provides the position or situation
of a region or entity. Often the locality of the region can become an
agent’s locality. Locality can be addressed in an absolute and
relational manner.
Absolute locators are locators that assign a unique address
to each agent or region. Simple two-dimensional grid system employ column-row
designations, geospatial systems employ longitude-latitude-altitude
designations, and IT systems employ unique identifiers in the form of
keys and unique reference IDs.
In contrast, relational locality means that an agent’s
location can be described as relative to another location. For instance
in a connected graph, one agent could be related to other agents, which
could in turn be related to other agents. Connected graphs such as the
web, electric power networks, or networks of colleagues are examples
of where entity’s location can be described relative to other
entities. In a planar environment an agent’s relative neighborhood
could be based on physical proximity rather than edges between nodes.
For example in a simple two-dimensional grid like a checkerboard, one
square can be characterized as diagonal to, or to the side of, and so
on. This kind of locality is particularly useful when an entity is constrained
to interact with the region of the environment that is near it. For
example, ant agents may only move from one region to an adjacent region,
and their pheromones might “flow” into neighboring regions
where pheromone strength lessens the further it travels.
Locality is useful for several reasons. One primary reason is that
communicating with an agent requires that the message can actually be
delivered to the agent. The sender of a communication may not be required
to know where the receiver is physically located, but at some point
the communication service must find the receiver to deliver the message.
Another reason is to provide location information. For example, a dispatcher
agent might need to know the physical location of its various resources
to schedule effectively. Lastly, agent movement or interaction may be
based on, or limited to, physical proximity. For example, an ant agent
may only move to or interact with the region that is immediately adjacent
to it. In contrast, a flea or grasshopper-style agent may jump multiple
squares in a single bound, but is limited to a maximum of five.
Fig. 6. Location-aware computing. [Buderi, 2001]
Also, it should be noted that since regions are positioned in an environment,
the notion of locality applies to them, as well. Furthermore, an agent’s
locality can be based on the locality of the region it occupies. For
example in the .geo example (Fig. 6), each region had an absolute locator
within which other entities are contained. In this way, an agent’s
locality can be defined in a discrete space, instead of locating the
agent one large, continuous environment.
Hybrid approaches using absolute and relative locality are also useful.
AT&T's "bat" transmitters are a good example of a hybrid
approach. "Bats" are small battery-powered ultrasonic transmitters
that can be worn on a belt or placed inside objects. They broadcast
a uniquely identifying 48-bit pulse to receivers embedded every 1.5
meters in ceilings as illustrated in Fig. 6. (For example, about 800
are placed around AT&T's three-story lab in Cambridge, England.)
Based on the known position of each receiver, the bearer's precise position
can be calculated. In other words, the transmitters and receivers have
absolute locators: the transmitter has an 48-bit ID and the receivers
are coordinate-based. Then, based on relative proximity, the coordinates
of the transmitter can be derived from the receivers’ coordinates.
Using this location information, zones can be established around objects
and people. If a person's zone overlaps an object's zone, the person
becomes the temporary owner of the device, be it a workstation, digital
camera, telephone or anything else. There is no logging on and everything
the user creates – documents, pictures, memos – is automatically
stored in the user’s personal files. [Baduri, 2001] In other words
this technology, known as location-aware computing, detects
when you're online and what kind of device you're using. Many companies
now have development efforts that involve location-based computing:
AT&T’s Sentient Computing R&D (described above), IBM’s
Pervasive Computing Division, HP’s CoolTown project, the ubiquitous-computing
projects at Intel and Xerox. [Want, 2001] Microsoft is another such
company with its new HailStorm services platform. When someone tries
to get in touch with you, the HailStorm system will detect your network
location and level of accessibility: Are you at your desk? In a meeting?
In transit? Depending upon the answer, the system will e-mail, page
or call you.
5 POSTLOGUE
By 2015, the social computing is expected by some to morph into ecological
or symbiotic computing. John Seely Brown, chief scientist of Xerox suggests
that structural matter (atoms) and computing (bits) will become inseparable.
Zillions of sensors, effectors and logical elements (made of
organic and inorganic materials) will be interconnected via wireless,
peer-to-peer technologies, producing smart, malleable stuff used to
build smart appliances, buildings, roads and more. It is during this
era that computers disappear. In their place, nearly every physical
artifact harbors some computationally based brainpower that helps
it know where it was, what was near it, when it was moved and so on.
In a way, the inorganic world took on organic properties, using computing
to transparently modulate responses to the environment. [Brown, 2001]
Footnotes
1 The action or influence of agents on each
other; i.e., having a reciprocal effect.
2 The agent communication channels
are defined as part of the physical environment. The communication environment
uses those channels to convey information.
3 Some debate exists about whether a single
agent can be its own group, because each agent can be thought of as
having both a social and physical existence. There is another debate
about whether or not a group has the status of an agent (holonics vs.
AALAADIN).
REFERENCES
[Brown01] Brown, John Seely, "Where Have All the Computers Gone?,"
Technology Review, 104:1, February 2001, 2001, pp. 86-87.
[Buderi01] Buderi, Robert (2001) "Computing Goes Everywhere,"
104:1, Technology Review, pp. 53-59.
[Castel00] Castelfranchi, Cristiano (2000) "Engineering Social
Order," Engineering Societies in the Agent World, Springer,
pp. 1-18.
[FIPA98] FIPA (1998) Foundation for Intelligent Physical Agents
FIPA98 Agent Management Specification, Geneva, Switzerland, Oct.
1998. (http://www.fipa.org)
[Odell00] Odell, James, H. Van Dyke Parunak, and Bernhard Bauer (2000a)
"Representing Agent Interaction Protocols in UML," First
International Workshop on Agent-Oriented Software Engineering 2000,
Paolo Ciancarini and Michael Wooldridge, eds., pp. 58-65.
[Odell00] Odell, James, H. Van Dyke Parunak, and Bernhard Bauer (2000b)
"Extending UML for Agents," Proc. of the Agent-Oriented
Information Systems Workshop at the 17th National Conference on Artificial
Intelligence, Gerd Wagner , Yves Lesperance, and Eric Yu, eds.,
Austin, TX, pp. 3-17.
[Parana01] Paranak, H. Van Dyke and James Odell (2001) "Representing
Social Structure using UML," Proc. of the Agent-Oriented Software
Engineering Workshop, Agents 2001 conference, Paolo Ciancarini
Michael Wooldridge, and Gerhard Weiss eds., Agents 2001conference, Montreal,
Canada, Springer.
[Pfeife99] Pfeifer, Rolf and Christian Scheier (1999) Understanding
Intelligence, MIT Press, Cambridge, MA.
[Russell95] Russell, Stuart and Peter Norvig (1995) Artificial
Intelligence: A Modern Approach, Prentice-Hall, NJ.
[SRI00] SRI (2000) http://www.icann.org/tlds/geo1/ModCE/E_TLD_POLICIES_FIN.htm
[Want01] Want, Roy and Bill Schilit (2001), “Expanding the Horizens
of Location-Aware Computing,” IEEE Computer, August 2001,
pp. 31-34.
[Weiss99] Weiss, Gerhard, ed. (1999) Multiagent Systems: A Modern
Approach to Distributed Artificial Intelligence, MIT Press, Cambridge,
MA.
About the authors
Cite this article as follows: James Odell, H. Van Dyke Parunak and
Mitchell Fleischer: "Modeling Agents and their Environment: The
Communication Environment", in Journal of Object Technology,
vol. 2, no. 3, May-June 2003, pp. 39-52. http://www.jot.fm/issues/issue_2003_05/column5
|
