Criteria for Building Interface Agents
Approaches to Building Interface Agents
Machine Learning Approach
Machine Learning Example
Using Intelligent Agents
Interoperation of Agents
Group Reasoning of Agents
Distributed Communicating Agents
How does an agent acquire the knowledge it needs to decide when to help the
user, what to help the user with and how to help the user?
How can we guarantee the user feels comfortable delegating tasks to an agent?
The rule-based approach features a collection of user-programmed rules for
processing information related to a particular task. An end user program
allows users to program rules for processing information for a task.
However, this approach has several problems. The agent acquires
its knowledge about how, when, and how much to help the user by extensive
programming by the user, which negates the purpose of an agent as a tool which
will save effort on the part of the user. Another drawback is that the
user needs needs to recognize the opportunity for using an agent,
program the rules and give the agent explicit knowledge, as well as maintain
these rules over time.
This interface agent has domain-specific background knowledge about the
application and user to recognize plans and contribute to users' tasks.
This approach has several problems as well. Where an end-user program
approach requires a great deal of work on the part of the user, this
approach requires a huge amount work for a knowledge engineer. The
knowledge engineer must outfit an interface with large amounts of knowledge
about the application, the domain and how the agent can help the user. The
knowledge of the agent is fixed and cannot be customized to individual users,
so its use by users may be limited. For example, in highly personalized
domains, the knowledge engineer cannot anticipate how best to aid the user.
This approach addresses problems encountered by the rule-based and
knowledge-engineered approaches. This approach requires less initial
work, and adapts over time. The agent acts as a personal assistant to
cooperate with a user on a task, but makes allowances for user override.
The agent learns by:
a) observing and imitating user
b) adapting based on user feedback
c) trained by user by example
d) ask for advice from other agents
Machine learning uses memory-based reasoning combined with rules to model
each user's habits. This approach achieves a level of personalizaton
impossible previously available except through user intervention. However,
these agents have their problems as well. Learning agents have a slow
learning curve, requiring a sufficient number of examples before it can
make accurate predictions. These agents also encounter problems when
dealing with completely new situations. To address these problems,
agents may learn from existing agents to get up to speed quickly.
Over time, agents learn to be selective when learning from other agents,
by learning to trust the suggestions of other agents more than others
for various classes of situations.
Machine learning agents apply several approaches to acquire knowledge.
Learning by demonstration involves continuously looking over the
shoulder of the user as the user performs actions. By monitoring the
user's actions over time, the agent finds recurrent patterns, and depending
on the agent's confidence in its assessment, the agent may offer to
automate the pattern of actions.
The confidence level of the agent is assessed by measuring situation
proximity by applying a weighted sum of the distance between
corresponding fields of two situations, then gathering the closest
matching situations in memory to calculate a prediction for an action
in the new situation. The agent's confidence in its prediction
is calculated by considering factors such as the number of situations in
its memory and how close the situations are to the present situation. The
user can set two confidence thresholds to control an agent's actions:
the tell-me threshold and the do-it threshold. An agent will make a
suggestion if the confidence level of an agent's prediction is higher than
the tell-me threshold. Likewise, the agent will autonomously take action
if the confidence level of the prediction is above the do-it threshold.
As an agent's experience base grows, the agent's confidence in its
predictions will grow. Thus, the user learns to trust the agent as the
agent becomes more 'trustworthy'. Monitoring facilities further reassure
the user of the accuracy of the actions of an agent. The activity monitor
is usually implemented as a caricature, which uses expressions to indicate
an agents state, such as alert, thinking, and working. The explanation
facility gives the user english descriptions of why an agent suggested
or performed an action. An interface to browse and edit the agent's
memory is useful to correct misconceptions on the agent's part.
First, independent agents must be implemented. Learning agents use
AI planning to express logical expressions of goals as input, then
generate a sequence of actions based on action schemata describing
available resources to achieve these goals, then synthesize and execute
plans to achieve goals.
Second, to take advantage of diverse agents with different abilities, there
is a growing demand for interoperation between agents. Integrating these
agents to exchange information and services with other programs independent
of individual agents' internal data structures and algorithms requires
a universal communication language to eliminate inconsistencies and
arbitrary notational variations. The procedural approach uses communication
as the exchange of procedural directives, which may require information about
the recipient that is not available to the sender, and cannot share information
both ways between sender and receiver. The declarative approaches uses
communication as the exchange of declarative statements such as definitions,
An agent communication language(ACL) consists of its vocabulary, an inner
language called KIF (Knowledge Interchange Format) and an outer language called
KQML (Knowledge Query and Manipulation Language). An ACL message is a KQML
expression whose arguments are sentences in KIF formed from ACL's vocabulary.
When agents have been enabled to interoperate using ACL, their collective
experience may be used to make decisions. Agents in multi-agent systems
perform different activities. Using group reasoning, agents deliberate
problems based on their current collective knowledge. Possible actions
are analyzed by communicating claims and arguing to support them
based on individual agents' experiences.
Agreement over an action is subject to dispute, which is settled by
feedback control systems, physical control systems and feedback phenomena.
Feedback control systems use conclusions pertaining to protocol and are led
back to the process by which they are generated. Physical control systems
use numerical analysis to understand what is going on. Feedback
phenomena relate back to human behaviour in making decisions to take action.
The beliefs held by individual agents dictate the arguments put forth,
and the outcome of disputes dictate what action to take, as well as
creating new beliefs.
DCA allows expert-systems to communicate with each other and with humans
to solve problems. Agents use models of each other and of the
resources available to them in order to cooperate, thus allowing for the
shared use of databases, knowledge bases, models of the business
environment, and process models.
DCA solves problems by using computational agents, each with their
own domain specific knowledge, and a number of users, assisted by an interface
agent or another agent to give problems to one of these computational agents
to solve. The computational agent proceeds to solve the problem by:
1) solving the problem itself using local knowledge and reasoning
2) decomposing the problem into subproblems, then
distributing subproblems to appropriate agent(s) and
integrating the subproblem solutions returned by these agents
When developing and using agents, the following issues should
Should information be filtered by agents?
New clips to a mailbox are still only accessible through that mailbox.
A mailbox with 100 items in it, even if the best of millions, are still
too many to be useful to someone with only 5 minutes to scan the news
before their next meeting. making folders can be as bad because it is
still just as inaccessible.
Should agents tell the truth?
If an agent can serve the purposes of their user by lying to less
intelligent agents, less intelligent agents may be manipulated into
serving the interests of a less truthful or more intelligent agent.
What will you trust agents to do?
Given your previous habits, you may allow an agent to decide for you
what you read, what kind of music you listen to, and who you interact
with. This may lead to a detailed although narrow view of the world,
instead of keeping you informed of a more general picture.
Should agents be merely slaves to its user?
Given that computers have the potential to learn so much more quickly
than humans, should we disable the full potential of our agents so that
agents will serve only our needs as opposed to their own interests?
Will we rely too much on agents, assuming they will do what we ask of them?
Back to report list