Compelling
Intelligent How
Larry
Birnbaum
Evanston
IL 60201
Redmond
Laboratory
20 Ames Cambridge
USA
Cambridge
+1 6172530338
(chair)
MA
impact
has not lived
appears to be changing. intelligent
use minimal whether
However,
user interfaces
or simplistic
have tended
Larry It’s
techniques
technology.
This
STATEMENTS
Intelligent
Content, systems
Silly! only perform
as well as their
work
repre-
can yield good performance. Permission to make digital/hard copies of all or part of this material for personal or classroom use is granted without fee provided that the copies are not made or distributed for profit or commercial advantage, the copyright notice, the title of the publication and its date appear, and notice is given that copyright is by permission of the ACM, Inc. To copy otherwise, to republish. to post on servers or to redistribute to lists. requires specific permission and/or fee. H-II 97, Orlando Florida USA . .$3,5(3 @1997 ACJf ()-89791.839..8/96/()1
propagation,
other
magic
bullet
In my view, the extent in a given
or
or tree search, or neural
application
is currently
to which these is mostly
a re-
presentation
will
describe
technologies
permedia
systems,
struction
tools,
our efforts
for building
educational
software,
and performance
support
to develop
structured
hy-
interface
con-
systems.
It
will focus on ASK Systems, a family of hypermedia case libraries developed at the Institute for the Learning Sciences to capture the ‘iwar stories” of human ex-
sentations of the task they are trying to perform and of the world they are trying to perform it in. If these representations are reasonable, then even extremely simple algorithms can result in useful performance. If they aren’t, then no algorithm, no matter how sophisticated,
in-
missed.
Birnbaum the
networks,
or whatever
such semantic POSITION
intelligent
flection of the extent to which a good job has been done in modeling the task and domain. However, by failing to focus on these modeling issues explicitly, an opportunity to build powerful semantic technologies is being
but some complex. Each panelist will then comment on the merits of different kinds and quantities of AI in the interface
– e.g., constraint
or belief
obsessing people.
IUIS.
of pragmatic
suggests that in building
algorithms
networks,
to
AI. In this panel we consider
The panelists will present examples of compelling IUIS that use a selection of AI techniques, mostly simple,
development
USA
1+ 4122687690
deduction,
more or less AI is the key to the development
of compelling
15213
Steven.
[email protected]
and not
This situation
so far the most inter(IUIS)
University
PA
terfaces, our main focus should be on content – the task being performed jointly by user and system, and the information being manipulated in this performance –
up to early expecta-
apparent.
Roth Institute
Mellon
Pittsburgh
This observation
of this work
esting
USA
markstlmerl.com
intelligence into the user interface for decades, but the commercial
and is not immediately
USA
.com
Steve Carnegie
02139
ABSTRACT Efforts to incorporate have been underway
98052
Robotics
+1 6176217534 .edu
WA
Way
djk@microsoft
201 Broadway
02139
[email protected]
tions,
Redmond
MERL
Street
MA
USA ,com
Joe Marks
Research
One Microsoft +1 2069362285
horvitz@microsoft
edu
Lieberman
Media
98052
Kurlander
Microsoft
Way
+12069362127
[email protected].
MIT
David
Research
WA
–
AI?
One Microsoft
USA
Interfaces
Horvitz
Microsoft
Avenue
1+ 8474913500
Henry
Much
Eric
ILS/Northwestern 1890 Maple
User
perts in video form, and make them available in a conversational mode of interaction implemented simply in graphical point and click interfaces. Over the past few years, dozens of these systems have been built, and many fielded, in domains as varied as military transportation planning, economic history, business-process reengineering, drug therapy for ALS, fluid-coupling design, and air-campaign commercial production poration. My talk will reflect joint Collins,
173
Roger Schank,
planning. They are currently in by the Learning Sciences Cor-
work with
Ray Bareiss,
and several other
Gregg
colleagues.
Eric
Horvitz
AI
and
the
On the
short-term User
Value
Interface:
le
of Inference
sequence of interface
where sophisticated
comes from
events.
inference
the domain
Another
may often
of monitoring
examp-
be crucial
applications.
It
Researchers attempting to enhance the computer-user interface have often pursued opportunities with the use of complex inferential machinery. Unfortunately, the sophisticated machinery frequently does not deliver great value, and can be mimicked by simpler techniques.
can be be critical to employ theoretically sound methods for managing the complexity of information displayed to a user in monitoring time-critical, high-stakes systems.
At times,
other form
inelegant
may even interfere In
the
context
application with
of automated
human-computer
of the
current
reasoning
of summarization.
I will highlight
interaction.
dominant
Such techniques must balance the costs and benefits of hiding information via pruning, abstraction, or use of
details
interaction
the value of inference
of Lumiere,
schemes, there is great opportunist y to make user inter-
by the Decision
faces more compelling
Microsoft
by focusing
effort on better
design
tion
on better
the possibility
procedures
UI design, taking
of integrating
into the functionality
face. We must remain
simple
the informational
into con-
There
automa-
creative rich
to use sophisti-
said, I would
also like to make a strong
case for
continuing, in parallel, research and development on more sophisticated user models and reasoning machinery for building More
compelling
sophisticated
intelligent
reasoning
cal and probabilistic
inference
can provide
means
complex
sequences
for detecting
grappling
with uncertainty,
for controlling faces. A
computer
tradeoffs system
learning
such M logi-
is often
faced
methods
der uncertainty
for reasoning about
Although
and
soft-
of design
automation
and
methods.
there
is also
understanding
of the
Kurlander More
the application
in building
uncertainty
making
with
Good of AI techniques
to user in-
terfaces still shows a great deal of promise, researchers in intelligent UI need to take a step back and gain perspective on the design tradeoffs that must be balanced
of user inter-
response
our
Than
for
and decision
user goals—in
work
by Lumiere,
to extend
Interface:
of events,
about a user’s goals and needs. Although good designs can minimize the uncertainty in a user’s intentions,
as demonstrated
Harm
Often
with
simple
in the
the
in the functioning
with
Intelligence
us with
usage patterns,
to be done in the realm
innovation
opportunity
David
user interfaces.
machinery
at
to reason about
role of sophisticated automated reasoning for enhancing the user interface, and, more generally, for enhancing the overall human–computer interaction experience. We need to continue vigorous work in both of these realms.
cated inference simply to get around poor design, or in lieu of better design combined with simple automation techniques. That
Systems Group
continues
needs of users.
is much
However,
of the user inter-
alert to attempts
and Adaptive
Lumiere
some
developed
ware by considering multiple user actions over time with Bayesian models. Lumiere also employs a probabilistic analysis of words in a user’s query and integrates the results from analyses of actions and words to identify
and straightforward pattern matching, similarity metrics, and search techniques. Overall, we can make great strides by focusing
Theory
Research.
system
the goals and needs of users as they
of layout, controls, and functionalist y of user interfaces, and, where necessary, weaving into the designs relatively straightforward automation such as basic event monitoring coupled with simple rules and control heuristics,
sideration
by describing
an experimental
real interfaces.
In designing
intelligent
the advantages
un-
to such
interfaces,
of using
efits of doing
things
terfaces
make mistakes,
often
it is critical
AI techniques
traditionally.
First,
to weigh
versus the benintelligent
in-
and the cost of verifying
evidence as a history of user actions—can be important in the operation of compelling interfaces. For a .vari-
and correcting
et y of tasks, it may be important to employ intelligentreasoning machinery for disambiguating the goals of the user, and determining the best action to take (i.e., selecting the action(s) that will maximize the expected utility of the user) in the face of inescapable uncertainty about a user’s goals.
erwise might be an interactive interface seem unresponsive. Third, users need a clear mental model of how the input, and some uses of computer will respond to their intelligence in the interface cloud this model. Fourth, users often want an explanation of why the system did something they way it did, and some forms of AI decision processes are difficult to convey to the users.
Consider the problem of guiding pelling dialogue with a user.
the generation of comEffective dialogue will
likely
processing,
hinge on natural
language
AI techniques
All
in conjunc-
inferences are often
of these disadvantages
can be prohibitive.
Second,
slow, and can make what
of using
oth-
AI in the interface
can be overcome. In many cases, it is a question of where to apply intelligence, so that the benefits outweigh the problems. In other cases, conventional interface techniques can be applied to support the use of
tion with methods that can make decisions given uncertainty about a user’s goals, the sense of words used by the user, the context at hand, and the long-term and
174
intelligence matic.
in the interface,
Ironically,
to make them interface
people
more
and make it actually often
natural,
apply
but
to support
the intelligence,
makes the interface
far less usable.
It is impossible technique
Suggest rather
prag-
AI to interfaces
without
crafting
the
or a new piece of interface
Much of the benefit of the Operate in real time. agent comes from acting while the user is ‘(busy”
●
Watch what the user is doing. Take advantage “free” information implicit in-user actions. -
of an interactive technology,
Steve Roth Iterating Between
both
Letting
and CHI
communities.
of
with-
out getting it in the hands of real people. The previous IUI conference was attended by a mix of people from the AAAI
act.
●
the AI component
to tell the true worth
than
UI researchers
the
AI
Product
and
UI
be the
Design:
Guide
already know (or should know!) that interface research is worthless without seeing how people truly respond to
It seems obvious to say that user interfaces must be judged by the ease and effectiveness with which they
the innovations.
are used by people to perform tasks. This means they usually need to be developed in concert with many other
A large
component
of UI research
is
empirical. Although I am not suggesting that all interface research needs formal user studies, it is still critical
facets of a product
- including
to make research systems available
other applications,
workspaces
so they
can see for themselves
interesting
to learn
being reported
for people to try out,
what
how many
works.
It will
of the research
at this conference
be
must
systems
are actually
be driven
and interfaces.
by a clear picture
with
UI design
of the product
pur-
pose.
available
Much of the research on intelligent
for people to try!
AI work, has been driven
Comic Chat is a graphical Internet I developed at Microsoft Research.
how they coordinate
chat program that It exploits some AI
gies, for example
interfaces,
by different
demonstrating
criteria
like other and strate-
the completeness
with
technologies in its interface, and is being used by thousands of people. I will briefly discuss some of what we
which a problem has been represented computationally or the correctness of an inference mechanism. Our experience in the SAGE project has been that when we
learned
shifted
uals,
from exposing
our research to so many individ-
and the effect of various
design
decisions
on the
attention
cused on usability
user experience.
needed
(sometimes Bit
Count
AI capabilities
user interface
in user interfaces,
so as not to
of the user. A key is to realize that many situations
user can be presented
are very underconstrained with
there may be no a priori another.
a wide variety
– the
of choices, and
reason for preferring
one over
In these situations, some AI software, often perceived by the user as an “agent” can help by intelligently making suggestions that help the user choose. I introduce the concept of “agent defaults” to describe this notion. Using
AI to help the user deal with
underconstrained
user interface situations casts the software in the role of a helpful assistant rather than an omniscient problem solver. I’ll
illustrate
this with
Letizia,
an agent that
helps the
user browse the Web. It acts as an “advance scout” recommending pages. Some key interface principles for successful integration ●
Don’t
&turb
of AI are: the
ways be possible
User’s
interaction.
It should
for the user to ignore
better
eliminating
changed direct the
dramatically.
manipulation need
and fo-
of problems
to
apply
that Some
techniques AI
tech-
as part of more complete product. I’ll illustrate these points with experiences in the design of automatic presentation systems.
as helpful or viewed as annoying. Some have suggested that the best approach is just to be very conservative risk the wrath
development
niques). Other problems created new interesting IUI research. The question is not how much AI or what kind. It’s how it can be selectively applied as needed
Incorporating artificial intelligence capabilities such as reasoning or learning into user interfaces can be viewed
in putting
product
issues, the kinds
to be addressed
changes required Henry Lieberman AI in User Interfaces: Making Every Little
towards
al-
the agent.
175