powered
by jeffsutherland.com/scrum
Object Management
Group
Framingham Corporate Center
492 Old Connecticut Path
Framingham, MA 01701-4568
U.S.A.
Telephone: +1-508-820 4300
Facsimile: +1-508-820 4303
Agent Technology
Green
Paper
Agent Working
Group
OMG Document ec/99-08-06
Version 0.8
20 August 1999
This paper presents a discussion of technology
issues considered in a Working Group of the Object Management Group The contents of this paper are presented to
create discussion in the computer industry on this topic; the contents of this
paper are not to be considered an adopted standard of any kind. This paper does
not represent the official position of the Object Management Group.
Table of Contents
1. Introduction...................................................................................................................................................................... 3
1.1 Purpose of the Green Paper................................................................................................................................. 3
1.2 Intended Audience(s)................................................................................................................................................ 3
1.3 Agent Working Group Mission........................................................................................................................... 3
1.4 Writing Team................................................................................................................................................................. 3
1.5 Questions & Comments............................................................................................................................................ 4
2. Current State.................................................................................................................................................................... 5
2.1 Definition of Agent................................................................................................................................................... 5
2.2 State of the Current Practice......................................................................................................................... 6
2.3 Maturity Level of the Industry..................................................................................................................... 6
2.4 Current Application Areas................................................................................................................................. 8
2.5 Agent Reference Architecture.................................................................................................................... 10
3. Key Areas of agent Technology...................................................................................................................... 11
3.1 Agent Communication.......................................................................................................................................... 11
3.2 Agent Mindspace....................................................................................................................................................... 12
3.3 Lifecycle Management........................................................................................................................................ 13
3.4 Agent as a Principle............................................................................................................................................... 16
3.5 Relationship between Agent and
Object Technologies.............................................................. 16
4. AGENT System Development................................................................................................................................. 21
4.1. Agent Modeling and
Specification............................................................................................................. 21
4.2. Testing, debugging,
validation, and simulation............................................................................. 21
4.3. Agent System development
methodologies....................................................................................... 21
4.4. Enterprise architecture and
Services.................................................................................................. 21
5. Considerations for Agent RFI/RFPs................................................................................................................. 25
5.1 Planned Roadmap..................................................................................................................................................... 25
6. Relationship to OMG Technology and Other Work Efforts................................................... 25
6.1. Current relevant OMG services.................................................................................................................. 25
6.2. Modifications and enhancements
of OMG specifications......................................................... 25
7. Other Similar Standards and Efforts......................................................................................................... 25
8. Appendix................................................................................................................................................................................ 26
8.1. Glossary........................................................................................................................................................................ 26
8.2. References................................................................................................................................................................... 26
8.3. Requirements for Agent
Technology...................................................................................................... 26
8.4. Open Issues.................................................................................................................................................................. 26
1. Introduction
1.1 Purpose of the Green Paper
Typically, when the word “agent” is used, many people conjure up images from James Bond to Web spiders. Yet, these same people feel that there is something more to the concept. For those who wish to know more, the purpose of this paper is to provide:
1) an overview of the general area known as agent technology
2) suggestions for standardizing areas of this new technology using the OMG process.
1.2 Intended Audience(s)
This green paper is intended for the OMG as a whole and, specifically, those task forces and working groups that are working in areas that might involve agent technology. This will ensure the work in these workspaces is well synchronized.
Another audience may be the agent technology industries outside of the OMG. Their thoughts and suggestions will help address the commercial significance of the concepts and directions established in the paper.
1.3 Agent Working Group Mission
The Agent Working Group of the Electronic Commerce
Task Force (ECTF) is to provide a forum for identifying and building consensus
and convergence in the industry around agent technology development.[1]
One of the first efforts of this group is to produce this green paper. This green paper has two purposes:
· to address the key features of agent technology, and
· to determine what and why the next steps might be for the Agent Working Group.
1.4 Writing Team
Editor: James Odell.
Contributors: David Kerr, Broadcom Eireann Research Ltd.
David Levine, IBM
Gregory Mack, Booz-Allen & Hamilton, Inc.
David Mattox, Mitre Corporation
Francis McCabe, Fujitsu Labs
Stephen McConnell, OSM sarl
James Odell
Kate Stout, Sun Microsystems, Inc.
Craig Thompson, Object Services and Consulting
1.5 Questions & Comments
Please address content-related questions or feedback to the Agent Working Group and to the paper editor James Odell (jodell@compuserve.com). In the interests of mailing list efficiency, please address errata or omissions directly to the editor only.
2. Current State
2.1. Definition of Agent
A basic dictionary definition of agent is one who acts. Under such a broad definition, agents can have a host of properties. One way to think about these properties is as follows:
· Autonomous - is capable of initiating action independent of any other entity; to some extent, can operate without direct intervention externally.
· Interactive - communicates with the environment and other agents, such as humans, machines, and software agents (i.e., social).
· Adaptive/reactive - responds to its environment.
· Mobile - able to transport itself from one environment to another.
· Proxy - may act on behalf of someone or something; i.e., acting in the interest of, as a representative of, or for the benefit of some entity.
· Rational - able to reason or understand.
· Proactive -
goal-oriented, purposeful; does not simply act in response to the environment.
· Intelligent - state is formalized by knowledge (i.e., beliefs, goals, plans, assumptions) and interacts with other agents using symbolic language.
· Temporally continuous - is a continuously running process.
· Character - believable personality and emotional state.
· Transparent and accountable - must be transparent when required, yet must provide a log of its activities upon demand.
· Able to learn and evolve - changes behavior based on experience.
· Cooperative - capabilities and resources can be exchanged among agents (e.g., can coordinate, collaborate, and negotiate).
· Rugged - able to deal with errors and incomplete data robustly.
· Trustworthy - adheres to Laws of Robotics and is truthful.
Most people who work with agents find this definition overly broad. However, a commonly agreed definition has not yet emerged. Some argue that an agent is not very useful without at least the first three of the above properties. The other properties listed above may or may not be present in agents to varying degrees. Rather than trying to determine if something is or isn't an "agent" it makes more sense to try and classify it along a spectrum that ranges from weak to strong agency. In general, the more of the above properties that an entity exhibits, the more agent-like it is. At a minimum, though, an agent is generally regarded to be an autonomous entity that can interact with its environment.
Software agents are a more specific type of agent. While there are many definitions, generally a software agent is an autonomous software entity that can interact with its environment. In other words, they are agents that are implemented using software. They are autonomous and can react with other entities, including humans, machines, and other software agents in various environments and platforms.
Agents are commonly considered similar, but not equivalent to objects Traditional object are biased towards notions, such as class, association, and message. While these constructs are useful for a certain category of applications, they do not directly address the requirements of agents. As presented above, agents have such characteristics as autonomy, mobility, and adaptability. Furthermore, business users like to express other concepts, such as rules, constraints, goals and objectives, and roles and responsibilities. (Stated another way, agents encapsulate the knowledge needed to perform a specific task.) In short, the agent-based approach distinguishes between autonomous, interactive, mobile objects (agents) and the passive objects of conventional OO. This does not mean that object orientation is dead; instead, it can be used to enable, rather than drive, agent-oriented technology.
2.2. State of the Current Practice
The emergence of agent technology is similar to the stories many other technologies (e.g. relational, object oriented, GUI). As such, some expect to some of history’s lessons to repeat themselves:[2]
· Agent technology is not a single, new, emerging technology—but rather the integrated application of multiple technologies.
· Agents are not necessarily a new, isolated form of application. Instead, it can add a new set of capabilities to existing applications.
· Initially, agent functions will emerge within applications, but later—with experience—will become part of the operating system or application environment.
· Agent applications may strengthen the human-computer interaction.
· Ultimately, applications that do not exploit agent support in the operating system will be severely disadvantaged.
While destined for lofty goals, the current state of agent technology is that:
· It is still an active research area.
· Isolated pioneer products are emerging.
· The full set of technologies are not available.
· The technologies are not integrated with one another.
· There is no consensus on how to support agents at the operating system level
· Despite the hype, agent technology is not in widespread use–nor has it been widely accepted as an inevitable trend.
· There are early adopters who can demonstrate the value of agent technology.
2.3. Maturity Level of the Industry
As suggested above, the agent “industry” is in an embryonic state. There are isolated examples of agents in use in industry. It is hard to pin down exactly due to the lack of a rigorous agent definition.
Because of this lack of singular definition, it is hard be specific about agents deployment in the software industry. However, there are several classes of agents that have been deployed to some degree.
· Network and system management agents
The telecommunications companies have been the most active in this area, and indeed seem the group most committed to the agent paradigm. Their goal is to use agents to assist in complex system and network management tasks, such as load balancing, failure anticipation, problem analysis and information synthesis.
One note: There is a confusing overlay of terms in this area. Network management is often performed by employing SNMP agents. These are a particular mechanism for management, but are not the types of agents we’re trying to discuss here. This is just a simple overlap of names.
• Decision and logistic support agents
Mostly deployed in closed environments, utility companies and military organizations use agents for information synthesis and decision support. These systems may alert an operator to a possible problem, provide information in support of a complex decision. They are closely aligned to decision support systems from the traditional AI community.
• Interest matching agents
These are probably the most used agents, and most users don’t even know they are using them. The interest matching agents are used by commercial Web sites to offer recommendations, such as
If you liked
“Frank Sinatra’s Greatest Hits” you might also like “Tony Bennett’s Songs for a
Summer Day”.
Based on Patti Maes work at MIT Media Labs, and later at Firefly, these agents observe patterns of interest and usage in order to make recommendations. They have been deployed at amazon.com, and various CD and video sales sites.
• User assistance agents
These agents operate at the UI level, offering information or advice to users. They are sometimes represented visually as a cartoon advisor. Companies such as Microsoft, Lotus, and Apple have shown the most interest in this area. The best known example of an agent in common use is the animated help characters used in Microsoft Office products. These agents use bayesean networks to analyze and predict possible topics that the user may need help with.
Each of these represent some aspect of agents’ features. However , none of these represent the full range of possible agent features.
The first commercial tool sets for building agents have entered the market within the last year. These agent building systems vary widely in functionality don’t adhere to any standards. Agents built in one system will not work in others. There is no uniform support for communications protocols across these tools either. The following table gives an overview of commercial agent systems available at the time of this writing.
|
Product |
Company |
Language |
Description |
|
AgentBuilder® |
Reticular Systems, Inc. |
Java |
Integrated Agent and Agency Development Environment |
|
AgenTalk |
NTT/Ishida |
LISP |
Multi-agent Coordination Protocols |
|
Agent Building Environment |
IBM |
C++, Java |
Environment |
|
Agent Development Environment |
Gensym |
Java |
Environment |
|
Agentx |
International Knowledge Systems |
Java |
Agent Development Environment |
|
Aglets |
IBM Japan |
Java |
Mobile Agents |
|
Concordia |
Mitsubishi Electric |
Java |
Mobile Agents |
|
Grasshopper |
IKV++ |
Java |
Mobile Agents |
|
iGEN |
CHI Systems |
C/C++ |
Cognitive Agent Toolkit |
|
Intelligent Agent Factory |
Bits & Pixels |
Java |
Agent Development Tool |
|
Intelligent Agent Library |
Bits & Pixels |
Java |
Agent Library |
|
JACK Intelligent Agents |
Agent Oriented Software Pty. Ltd. |
JACK Agent Language |
Environment |
|
Jumping Beans
Engineering |
Ad Astra |
Java |
Mobile Agents |
|
Kafka |
Fujitsu |
Java |
Agent Library |
|
LiveAgent |
AgentSoft Ltd. |
Java |
Internet Agent Construction |
|
Microsoft Agent |
Microsoft Corporation |
Active X |
Interface creatures |
|
Odyssey |
General Magic |
Java |
Mobile Agents |
|
Via: Versatile Intelligent Agents |
Kinetoscope |
Java |
Agent Building Blocks |
|
Voyager |
Object Space |
Java |
Agent-Enhanced ORB |
2.4. Application of Agents
2.4.1. Enterprise applications
· Smart documents (i.e., documents that `know' that that they are supposed to be processed)
· Goal-oriented enterprise (i.e., work-flow on steroids)
· Role and personnel management (i.e., dynamically attaching roles and capabilities to people)
2.4.2. Inter-Enterprise applications
· Market making for goods and services
· Brokering of the above
· Team management
2.4.3. Process control
· Intelligent buildings (e.g., smart heating/cooling, smart security)
· Plant management (e.g., refinery)
· Robots
2.4.4. Personal agents
· Email and news filters
· Personal schedule mgt
· Personal automatic secretary
2.4.5. Information management tasks
· Searching for information - The amount of information available over a corporate intranet stretches the capability of most users to effectively retrieve useful information. Search agents contain domain knowledge about various information sources. This knowledge includes the types of information available at each source, how to access that information and other potentially useful knowledge such as the reliability and accuracy of the information source. Search agents use this knowledge to accomplish specific search tasks..
· Information Filtering - Another common task for agents. Information filtering agents attempt to deal with the problem of information overload by either limiting or sorting the information coming to a user. The basic idea is to develop an on-line surrogate for a user that has enough knowledge about the user's information needs so that it can select only those documents that would be of interest. These types of agents usually function as gatekeepers by preventing the user from being overwhelmed by a flood of incoming information. Filtering agents also work in conjunction with, or are sometimes incorporated into, search agents in order to keep the results from searches down to reasonable levels. Typically, filtering agents incorporate machine learning mechanisms. This allows them to adapt to the needs of each user and to provide more precision than that typically provided by keyword filtering approaches.
· Information Monitoring - Many tasks are dependent on the timely notification of changes in different data sources. A logistics planner my develop a plan for moving equipment from one location to another, but the execution of that plan could be disrupted by the onset of bad weather at a refueling stop. The logistics planner would like to know of any events that would be likely to effect his plan as soon as they happen. Agents are useful for monitoring distributed data sources for specific data. Being software constructs, they have the patience necessary to constantly monitor data sources for changes. Alternately, mobile agents can be dispatched to remote or otherwise inaccessible locations to monitor data that the user might not normally have access to.
· Data Source Mediation - The data management landscape is populated with a multitude of different systems, most of which don't talk to each other. Agents can be used as mediators between these various data sources, providing the mechanisms that allow them to interoperate. The SIMS Project at ISI developed an information mediator that provides access to heterogeneous data and knowledge bases. This mediator can be used to create a network of information gathering agents, each of which has access to one or more information sources. These agents use a higher level language, a communications protocol, and domain-specific ontologies for describing the data contained in their information sources. This allows each agent to communicate with the others at a higher semantic level.
· Interface Agents / Personal Assistants - An interface agent is a program that is able to operate within a user interface and actively assist the user in operating the interface and manipulating the underlying system. An interface agent is able to intercept the input from the user, examine it, and take appropriate action. While interface agents are not directly related to data management, they have to potential to play a large role in assisting users of data management systems. This becomes increasingly important as data management systems become more distributed and group together to form large, complex systems of systems. Agents in the interface can function as a bridge between domain knowledge about the data management systems and the user. These agents could assist users in forming queries, finding the location of data, explaining the semantics of the data among other tasks. Examples of this include intelligent tutoring systems and web browsing assistants [[3]]. In addition, Microsoft is now including interface agents in its desktop products to watch the actions of users and make appropriate suggestions.
2.5 Agent Reference Architecture
Craig Thompson has submitted
this as a possible taxonomy. Can it be
integrated in the paper?
3. Key Areas of Agent Technology
Several key areas should be considered for systems that employ agent-based technology.
3.1 Agent Communication
Currently, one of the most important areas for standardization is agent communication. If every designer developed a different means of communicating between agents, our agent systems would be worse than a tower of Babel. Not only would the content and meaning of a communication likely be different, but the means of communication could occur in a variety of ways.
3.1a Agent communication languages
Messages must have a well defined
semantics that is computational and visible.
Thus, we need standardized agent communication languages (ACL), so that
different parties can build their agents to interoperate. Furthermore, they must have a formal
semantics so that different implementations preserve the essential features of
the ACL. By specifying an ACL, we
effectively codify the basic elements of interaction that can take place
between agents.
Possible
implementations:
KQML
Arcol and FIPA
KIF
XML-based
3.1b Message transportation mechanism
In agent environments, messages should be schedulable, as well as event driven. They can be sent in synchronous or asynchronous modes. Furthermore, the transportation mechanism should support unique addressing as well as role-based addresses (i.e., “white page” versus “yellow page” addressing). Lastly, the transportation mechanism must support unicast, multicast, and broadcast modes and such services as broadcast behavior, non-repudiation of messages, and logging.
Possible
implementations:
CORBA
message = structured event as per CosNotification services
JAVA messaging service
RMI
DCOM
Enterprise Java Beans Events
3.1c Ontology communication
Agent communication implies that concepts will be part of a communication among agents. Furthermore, agents can have different terms for the same concept, identical terms for different concepts, and different class systems. A common ontology, then, is required for representing the knowledge from various domains of discourse. The two primary challenges here are building them and linking them.
Possible
implementations:
UML
MOF
OKBC
XML DTD
3.1d Agent protocol
Agents can interact in various patterns. Cooperative agents work toward a common goal. For example, to produce a coherent plan, agents must be able to recognize subgoal interactions and either avoid or resolve them. Partial Global Planning does not assume any distribution of sub-problems but allows the agents to coordinate themselves. Joint intention frameworks require joint “commitment.” The Shared Plan model is based on the mental attitude—the intention to act. Joint Responsibility is based on a joint commitment to the team’s goal and the recipe for attaining that goal. Self-interested multiagent interactions are usually centered around negotiation and contract nets. Whichever kind of conversation is chosen, the pattern—or protocol—of that conversation must be understood by the participating agents. Each protocol is a pattern of interaction that is formally defined and abstracted away from any particular sequence of execution steps. Standards and guidelines here would greatly accelerate our usage of agents, because they could then “converse” without requiring human intervention.
Possible
implementations:
Negotiation RFP
3.2 Agent Mindspace
As we design and build agents that are more intelligent, we need to consider aspects such as understanding, adaptation, beliefs, desires, intents, and knowledge.
3.2a Goal representation and manipulation
Many theories of goals exist. In the early work on problem solving, a goal was a state to be achieved or a proposition to be made true. However, agent goals could be seen as divided into various classes: achievement goals (long-term goals), satisfaction goals (recurrent goals such as resource gathering), preservation goals (for perserving life and property), and delta goals (i.e., “other” state changes). Also, in the BDI (beliefs, desires, intents) approach, “desires” describe the agent’s goals (sometimes including a motivational aspect) and “intentions” characterize the goals (desires) that the agent has selected to work on. Goals can have sub-goals, particularly in problem solving. We need to address both the maintenance and the communication of goals.
Possible
implementations:
CosTrader constraint language and the CosNotification language extensions
OCL
XML-based
3.2b Reactivity versus proactivity
One of the minimum requirements for an agent is that it is reactive. Reactive agents can selectively sense/perceive events (state changes, messages) in their environment and respond to them. The syntax for reactive agents is typically in the form of WHEN event IF condition THEN action/assertion. As such, inference engines can be used by reactive agents. In contrast, proactive agents do not simply react, they should be able to exhibit opportunistic, goal-directed behavior and where appropriatetake the initiative. Proactive agents can choose to query the environment about its state rather than just wait until events arrive; that is, they are active rather than just receptive. Agents that continuously reason about future events are sometimes called deliberative.
3.2c Adaptive agents
Adaptive agents can be simply reactive or they can learn or evolve. Agents that learn or evolve can change their behavior based on their experience with other agents and the environment. Here, learning and evolution is orthogonal to reaction/proaction.
3.2d Procedural versus declarative process specifications
Procedural approaches specify how a computation should proceed; declarative approaches specify what an operation should do rather than how. Both approaches have their merits and therefore both should be considered.
Possible
implementations:
UML Activity Diagrams
UML process specification language
Workflow RFP
3.2e Comprehension of an environment
In order for an agent to recognize
features and characteristics of an environment, interfaces are required through
which an agent can obtain service information.
The CORBA 2.2 specification
provides the get_service interface (PIDL) at the level of the ORB. An equivalent interface is required to
support the registration and retrieval of information concerning available
facilities and applicable policies. Such interfaces should enable an agent both
to evaluate an environment and to register itself with the environment as a
potential services provider.
3.3. Life-Cycle Management
Agents will be software running in software environments. As such, they must have well understood mechanisms for such activities as starting, stopping, being managed, and being traced. Some agents are implemented as mobile code, which introduces additional lifecycle issues, such as permissions to run, permissions to perform certain tasks, and to have communication occur at different locations other than its original starting point. Finally agents can evolve and can possibly clone themselves, which introduces issues related to the delegation of responsibilities and permissions.
As we examine lifecycle management, we must also examine the software environments in which agents will run. These can be very small, intermittently connected devices like cell phones or Personal Digital Assistants, to very large clusters of servers capable of running huge numbers of agents. The requirements for each environment are quite different, and each must be accommodated.
3.3a Virtual existence or persistence
Agents may logically exist for indeterminate periods of time during which they may display dormant behavior. Interfaces supporting agent lifecycles need to consider requirements for “logical” existence of possibly very large numbers of agents as opposed to in-memory physical existence. This has a number of implications around storage, messaging and communication, and management.
• Persistence of agents
In systems that currently support long-lived agents, agents can be “put to sleep” and saved (usually to a physical disk, although this is just an implementation strategy). When the agent is persistent, the current state and data are preserved, and are restored when the agent is “awakened.” There are often other state transitions associated with “putting the agent to sleep” and “awakening” it, having to do with querying whether it is currently in a situation where this can successfully be performed, and on awaking, a notification that time has elapsed and it might want to reassess the environment.
• Messaging while “asleep”
A variety of messaging models are possible for agent systems. Some may need a full robust event-style model with guaranteed message delivery. Others may need a far more casual model where an agent seeing a message of interest may act on the information. There may be messaging models for direct communications between two agents, for a store-and-forward approach, and for a publish-and-subscribe model. System events may also be passed via messages or some other mechanism. Design decisions about what intersection of these strategies to use depend on the type of agent applications being built.
These design decisions also impact how messages are handled while an agent is “asleep.” If an agent has to be awakened with each incoming message to see if the message is relevant, storing the agent to save running many in-memory copies would be counterproductive. Models must be designed to optimize delivery of messages. There will likely be many different models for this, but over time certain design patterns will emerge.
• Managing agents
Agents that are currently “asleep” also need to be managed. In large distributed systems, agents could be moved to a new server in order to provide dynamic load balancing.
3.3b History
Mechanisms are required to provide
a historical recording of the agent’s actions—so that agent behavior can be
audited and that agents can evaluate prior actions. This could include a range of situations,
from merely obtaining an agent's state to providing a comprehensive log of
actions.
3.3c Mobility
Static agents exist as a single process on one host computer; mobile agents can pick up and move their code to a new host where they resume executing. From a conceptual standpoint, such mobile agents can also be regarded as itinerant, dynamic, wandering, roaming, or migrant. The rationale for mobility is the improved performance that can be achieved by moving the agent closer to the services available on the new host.
Mobile agents can also be part of an agent system that has static agents. For example, there may be a large “intelligent” agent coordinating a set of actions (such as a workflow, a negotiation, or data synthesis), which in turn sends out smaller mobile agents to perform a specific task, such as acquiring some data.
Mobile agents create an additional set of requirements:
· They require an agent server where they can run.
· They introduce the complexities of security and validating the authenticity of mobile code.
· They introduce management complexity on such issues as:
* Where are they?
* How do you communicate with them when they are moving about?
* How do you bring them home when network failures occur?
* How do you ensure that they have a reasonable way of “timing out” and halting a process if the task is time sensitive?
* If they must keep functioning despite adversity, how do you ensure they stay alive?
· Naming and identification become very important.
3.3d Process centric
If an agent is primarily thought of
as a process, you can observe and view the state of the agent. For such agents, it is appropriate to
establish relationships between agents that consume and agents that
produce. For example, an agent version
of Task Session, you can describe resource agents, task agents, and process
agents. (This one needs more
explanation)
Possible implementations:
Virtual Existence: POA servant management provides explicit support for the management of many logical instances
History: event logging
Mobility: extensions to the CosLifeCycle move() operation and associated POA virtual existence management.
3.3e Dynamic and multiple classification
During an agent’s lifecycle, the interfaces exposed by the agent may be dynamic, reflecting changes in its state or environment. Mechanisms are required to support the existence of multiple interfaces relative to a single agent identity. Furthermore, control mechanisms must be established to enable and disable an agent’s features. Such mechanisms are particularly useful when multiple roles are being supported. Depending on the situation, the roles or behavior of individual agents can change. For example in ant societies, workers can become foragers when food availability increases, and nest-maintenance workers can become patrollers when foreign ants intrude. However, once an ant is allocated to a task outside the nest, it never returns. This implies that an agent can both change its role (dynamic classification) and assume more than one role at any moment in time (multiple classification).
Possible implementations:
interface variation dependent on state
CORBA Component equivalence
3.4. Agent is a Principal
A key attribute of an agent is that it be able to act autonomously. Agents can then take on a wide range of responsibilities on our behalf, including purchasing goods and services and entering into agreements. Furthermore, agents may have several identities which they can use while operating. Any robust system will consider these identities in its transactions with an agent.
3.4a Agents are operating on behalf of an entity
An agent will often perform some tasks on behalf of another. For example, a software agent could perform a task on behalf of a person. It could also perform on behalf of another piece of software (another agent), an organization, or a particular role (manager, system administrator). This leads to an number of issues that need to be considered:
· An agent needs to be able to identify on whose behalf it is running.
· If it requests an action from another entity, the other entity may want to assess whether it will permit the action, based on the proffered identity. This might include assessing things such as:
* Knowledge of the identity.
* Some set of permissions that control what this identity is permitted to do.
* Whether the identity has a good credit (in a financial transaction) or a good reputation.
* Results of previous interactions with that identity.
Subsets of permissions are another issue. In human interactions, we delegate portions of our decision making routinely as well as in an ad hoc manner. For example, we might ask one of our co-workers the weekly status meetings staff meeting, knowing that the person will pick a convenient time, include the right people, and schedule an appropriate room. To another co-worker, we may delegate signature authority to spend up to $1000 in our absence. In both cases, co-workers have been delegated authority but given very different permissions. A similar set of constructs needs to be created for agents.
3.4b Agents are software
Agents are software and as such may be more or less reliable. Various schemes permit unknown software to run in an environment. For example, code signing might be employed to determine whether the agent will be permitted to run in a given environment. Since an agent may in fact be composed of multiple objects, a model may be needed to assess the overall reliability of all components.
3.5. Relationship between Agent and Object Technologies
As the Object Management Group primarily concerns itself with the interoperation of object systems, and the Agents Working Group is part of OMG, it is reasonable to examine the relationship between agents and objects. In particular, it is tempting to assert "An agent is an object that...", completing the phrase in a variety of ways. We will argue that such definitions miss key features of agency. We will then discuss why, if Agents are arguably not objects, the OMG should work to provide standards for agents.
For this discussion, we will focus on the key features of agents defined in chapter 2, namely that agents are characterized by their Autonomy, their ability to Communicate and their reactivity, or ability to Act. We will discuss how these features contribute to describing the behavior of an Agent.
In describing the key features of objects, we will begin with Grady Booch's triple of State, Behavior and Identity. We will then observe that in an OMG context, behavior is largely characterized by IDL, or more generically by the methods available to access and act on the object.
As a final bit of stage setting, we note that OMG, and in particular the OMA, extends beyond the mere support of basic objects, and includes both Object Services, the Common Facilities and Domain interfaces.
3.5.1 Modeling an agent as an object
When we model any software entity as an object, we focus on its interfaces and its state. In general, the single key characterization of an agent's behavior is that it accepts messages from other agents, performs some processing and (possibly) responds. This behavior, in terms of interface, may be represented in two forms. One would be to provide a method for every message that the agent is capable of processing. The other would be to provide a single method, "AcceptCommunicativeString" which would permit the agent to accept arbitrary messages.
One method per message
At first, it is tempting to imagine providing an agent with a method for each message it can accept. One would build a class hierarchy of methods and use classic OO design patterns to structure these methods into classes, specializations, and eventually individual methods. However, there are several issue that arise in practice.
There are expressive limitations in a scheme that maps communications directly to method calls, and there are issues pertaining to autonomy, and how to model agents that engage in multiple, long lasting conversations.
• Expressive limitations
The expressive limitations of modeling agent messages as method calls are several in nature. By using method invocation to express part of the message (in essence the performative) the agent no longer has direct reference to the entire message being sent to it. This also limits the types of messages that may be sent to an agent to those for that messages exists, or it requires splitting the world into two portions, namely those messages for which we have methods, and those messages that we send as strings, using an "AcceptCommunicativeString." method. Since we may wish to send any message to any agent, and we want the expressive power to cover all these cases, we in fact must be able to express every communicative string, even ones for which methods exist, in communicative language.
• Autonomy
Method invocation also has the wrong basic connotations in terms of autonomy. In general, method invocation is done either on the caller's thread of control, or on the thread of a remote proxy for the caller, in the case of remote method invocation. Agents, on the other hand, aren't running as part of the caller, nor are they running with the caller's identity and privileges. In fact, quite the contrary, an agent runs on its own behalf, and with its own identity and privileges While one can imagine doing things with access control such that the additional rights and proper identity were associated with each invocation, it adds steps that are outside the normal notions of method invocation.
Beyond identity and access control rights, we also run into the notion of autonomy. Unlike a normal object, where access control rights and visibility largely encapsulate notions of whether a method should run, an agent is an autonomous element of the software world. Thus, for every message that the agent receives, it may determine, based on its own goals and state, and the state of the ongoing conversation, whether to process the message and how to respond if it does. This implies that every method, will, in effect, begin with the idiom of making such a determination, something which, in general, is beyond what the majority of objects do at method invocation time. (Parameter checking, and comparison to internal state is fairly normal, but knowing the identity of one's caller, and basing decisions on that, is not)
• Conversations, and long term associations
Another problematic issue for thinking of agent messages as method invocation is the question of long term conversations and associations. Agents, unlike simple objects, may engage in multiple transactions concurrently, either through the use of multiple threads, or similar mechanisms. This is expressed naturally in a messaging environment, as each conversation can be assigned a separate identity, and either a unique message destination or a unique identifier can be used to sort out the threads of discourse.
Similar mechanisms can be built out of objects, of course, but they would bypass, or be in addition to, normal method invocation. One could imagine producing multiple cloned "copies" of the agent object and passing one reference to each counterpart, so as to provide separate communication contexts for each conversation. But the very need to invent this kind of mechanism implies that the agent isn't acting as a simple object.
Action
A second set of issues obtains when we consider that agents are active, not passive entities within a software system. Traditionally, an object is passive, with its methods being invoked under the control of another component of the system. An agent, in general, can react not merely to specific method invocations, but to observable events within the environment. Further, as part of the basic structure of multi-agent systems, it is natural for agents to be engaged in multiple, parallel interactions with other agents.
Public Interpretation
Communicating agents, generally express their messages in a way that allows a public interpretation of the message traffic (visibility subject to access control and other issues) This adds significant expressive power, at the cost of requiring explicit onotologies. This expressive power is rather different from the normal notion of methods, and of class hierarchy and the interpretation of the meaning of a method.
Scale of agents
While there isn't any single grain size that characterizes all agents and multi-agent systems, agents can often be larger than single objects. While it is sometimes entirely reasonable to model large, complex entities as objects, the structure and sorts of parts that comprise agents reduce the effectiveness of thinking of agents as simple, unitary objects. For example, many agents include complex reasoning components, such as inference engines or neural networks. These resources are often implemented as sharable sub-systems, used by many agents, or used as parts of platforms supporting many agents.
It should be added that, while agents can any size, development guidelines often suggest constructing agents that are small in mass, size, and time.
• Small in Mass
In systems with large numbers of agents, each agent is small in comparison with the whole system. For example, each termite is an almost negligible part of the entire termite hill. As a result, the behavior of the whole is stable under variations in the performance of any single member, and the collective dynamics dominate.
• Small in Time (Forgetful)
Naturally occurring agent systems can forget. Pheromones evaporate, and as a result obsolete paths leading to depleted food sources disappear rather than misleading members of the colony. Even the death of unsuccessful organisms in an ecosystem is an important mechanism for freeing up resources so that better adapted organisms can flourish. Furthermore, each transaction can cost and agent, leading to a decrease in energy over time.
• Small in Scope (Local Sensing and Action)
The participants in natural systems usually can sense only their immediate vicinity. In spite of this restriction, they can generate effects that extend far beyond their own limits, such as networks of ant paths or termite mounds. Wolves are built close to the ground and so can’t see very far, but manage to coordinate a hunt with other wolves over a wide area.
3.5.2 The single method agent
Given the above arguments, one could contemplate creating an agent class, which, in essence had one, and only one primary method, namely "Accept Communicative Message." Such a construct is certainly possible, but at that point, you've lost any descriptive power from the object model, and you've reduced environments such as CORBA to playing the role of basic message transport.
3.5.3 Agents and agent systems composed from objects
While we make the case that agents are not the same as objects, this doesn't mean that agents shouldn't be composed out of objects. In fact, as with any larger software component, we strongly advocate composing agents from objects, and, indeed building the infrastructure for agent based systems on top of the kind of support systems usd for complex object-oriented software systems
Agent Components
When we begin to decompose agent structure, we rapidly find lots of structures and parts that are reasonable expressed as objects. These might include:
• Agent Names
• Agent Communication handles
• Agent Communication Language Components, including:
- Encodings
- Ontologies
- Vocabulary Elements
• Conversation Policies
Agent Infrastructure
Beyond the parts that make up agents, there is an additional layer of software components in multi-agent systems that may be naturally expressed as objects and collections of objects. This is the underlying infrastructure that embodies the support for agents composed of object parts.
Examples of such might include:
• Communication Factories
- Transport References
- Transport Policies
• Directory elements
• Agent Factories
3.5.4 Relating agents to OMG
Positioning OMG as the obvious integrating technology
OMG, through CORBA, the OMA framework and its related services, provides almost all of the elements needed to produce a rich framework for supporting agents. Further, through IDL, ORBs and IIOP, OMG provides an excellent basis for knitting together abstract agent frameworks instantiated in multiple OMG compatible environments.
Java, with its strong ties to OMG, provides an excellent basis for building Agent Frameworks. Indeed, many agent projects are being built out of Java parts, including Java Beans, RMI, and IIOP. OMG should play a major role in bringing these elements together, and standardizing the usage of various OMG technologies in these areas so as to enhance interoperation and portability.
OMG would need to provide standards in a number of areas, if it desires to act as a major player in the agent arena. Ontologies Agent Communication Languages and services for managing these elements would be clear candidates. Infrastructural elments, such as transport factories, agent support platforms and the like would also fall into this purview.
Integrating Agents in the OMG world view
A more comprehensive proposal would entail integrating agents into the OMG world view as "first class" elements in the overall set of OMG abstractions. This would include supporting the ability of agents and objects to name, invoke and manage each other as first class software elements. Such a proposal would represent a fundamental revisiting of many basic assumptions in the current OMG perspective on software. Such a proposal does not seem warranted today, with the current state of agent technology, and current marketplace acceptance of agent-based systems. However, as the state of the art advances, the Agent Working Group should keep this option available, and, in particular, avoid taking steps that would preclude tighter integration of Agent technologies into the OMG core in the future.
4. Agent System Development
4.1 Agent modeling and specification
We need to better understand how to develop applications using agent technology. Therefore, development tools and methods play an important part in agent-based systems. Here, we need to determine what kinds of modeling languages are required as well as the underlying metamodels needed to support agent specification and deployment. Furthermore, we need to ensure interoperability across the lifecycle of agent tools and their designs/work products.
Possible implementations:
UML
MOF
4.2. Testing, debugging, validation, and simulation
Concurrency and nondeterminism in execution make it very difficult to comprehend the activity of an agent system. We need visualization tools, debugging facilities, and simulation.
4.3. Agent system development methodologies
We must determine the steps needed to guide the life cycle of agent system development—at both micro and macro levels. Existing methods can be used to some extent, but we need to identify how an agent-based approach will change them.
Figure 1. Agent Management Reference Model-
4.4. Agent enterprise architecture and services
For the most part, agents will be deployed within conventional enterprises and will draw on the enterprise for many services. CORBA provides a rich source of services and a proven architecture. This section provides a framework for considering how a system supporting agents might draw on CORBA services and facilities. The architectural basis for this discussion will be the FIPA architecture.[4] The FIPA Agent Platform provides a good construct from which to discuss the enterprise-related issues in agent deployment. Figure 1 is a UML depiction of the FIPA98 Agent Management Reference Model that has been augmented to include components to facilitate interfaces to non-agent software and services.
4.4a Agent Platform
The key element to the enterprise
architecture is the Agent Platform. An
Agent Platform (AP) provides an infrastructure in which agents can be
deployed. An agent must be registered on
a platform in order to interact with other agents on that or other
platforms. Minimally, an AP consists of
three capability sets: an Agent
Communication Channel, an Agent Management System, and a Directory
Facilitator. The Internal
Platform Message Transport (IPMT) is the local (possibly proprietary) means of
exchanging messages within an AP. Figure 1 is a UML depiction of the FIPA98
Agent Platform Architecture.
Figure 2. Agent Platform Architecture
FIPA does not specify the physical nature of a platform. However, two cases should be considered, that of a single host and that of multiple processors deployed as a “virtual” platform. If the Platform is virtual, having it fulfill several requirements would be wise. It should have:
1. high-speed communications.
2. a single system manager.
3. a single security enclave.
These last two requirements make the agent system easier to use. The responsibility for humans managing the system is simplified. Also, we avoid the situation where control of the agent lifecycles on a platform is shared by several people. From the system perspective, the lifecycle of all agents in a given platform is controlled by a single entity—the Agent Management System. From the perspective of a human (or agent proxy), the platform itself should also be controlled by a single entity.
4.4b Agent Management System
The Agent Management System (AMS) is an agent that supervises access to and use of the AP. Only one AMS will exist in a single AP. The AMS maintains a directory of logical agent names and their associated transport addresses for an agent platform. The AMS offers “white pages” services to other agents. The AMS is responsible for managing the lifecycle of the agents on the platform.[5] Its actions include:
1. Authentication
2. Registration
3. De-registration
4. Modification
5. Query platform profile
6. Search
7. Control of agent lifecycle
4.4c Directory Facilitator
The Directory Facilitator (DF) provides “yellow pages” services to other agents. The DF is a mandatory, normative agent. Agents may register their services with the DF or query the DF to find out what services are offered by other agents.
4.4d Agent Platform Security Manager
The Agent Platform Security Manager (APSM) is responsible for maintaining security policies for the platform and infrastructure. The APSM is responsible for run-time activities, such as communications, transport-level security, and audit trails. FIPA 98 security cannot be guaranteed unless, at a minimum, all communication between agents is carried out through the APSM.
The APSM is responsible for negotiating the requested inter- and intra-domain security services with other APSMs in concert with the implemented distributed computing architecture, such as CORBA, COM, and DCE, on behalf of the agents in its domain. The APSM is responsible for enforcing the security policy of its domain and can, at its discretion, upgrade the level of security requested by an agent. The APSM cannot downgrade the level of services requested by an agent but must inform the agent that the service level requested cannot be provided.
4.4e Agent Resource Broker
An Agent Request Broker (ARB) is an agent that brokers a set of software/service descriptions to interested agents. Clients query it about what non-agent software services are available.
4.4f Wrapper Agent
This agent allows an agent to connect to a non-agent software system/service uniquely identified by a software description. Client agents can relay commands to the wrapper agent and have them invoked on the underlying services. The role provided by the wrapper agent provides a single generic way for agents to interact with non-agent software systems.
The wrapper increases robustness by exposing an agent interface capable of exception and event handling for the requester. The use of an adapter also allows the agent manager to control platform access to the remote service.
4.4g Agent Communication Channel
All agents have access to the Agent Communication Channel. It provides a path for basic interchange between agents, agent services, AMS, and other agent platforms. At least, it must support IIOP. Agents can reach agents on any number of other platforms through the Agent Communication Channel. Access to agents outside of the local namespace could be supported by the CORBA Trader Services.
5. Considerations for Agent RFI/RFPs
It is anticipated that RFIs and/or RFPs will be issued for agent-based technologies. For example, the overall goal of RFI-1 is to collect information regarding agent technology from various communities. This information will help guide the OMG in the adoption of specifications that will to extend the OMG Object Management Architecture with agent technology functionality. These extension, then, would further populate or align the OMG with emerging agent standards, protocols, tools, and utilities. Currently, there are no other RFIs or RFPs being considered or in development.
5.1 Planned Roadmap
A possible plan of attack
is to see how this coordinates with FIPA, and to figure out what might make
sense within the OMG domain. For example, I can’t really picture User
Assistance (Desktop) agents as being with the scope of OMG.
FIPA did approve doing an
architecture that could map to various implementations, such as CORBA.
6. Relationship to OMG Technology and Other Work Efforts
6.1. Current relevant OMG services
tk
6.2. Modifications and enhancements of OMG specifications
tk
7. Other Similar Standards and Efforts
FIPA
KQML/KIF
US DARPA
CoABS -
Jim Hendler <jhendler@darpa.mil
ALP -
Todd Carrico <tcarrico@darpa.mil
EU AgentLink community
other
8. Appendix
8.1. Glossary
tk
8.2. References
tk
8.3. Requirements for Agent Technology
tk
8.4. Open Issues
tk