Projects







Knowledge Bases:



    
  
  • Construction of semantically rich models: LOOM
  
  • Tools to organize and check model consistency: LOOM
  
  • Tools for merging large models: SENSUS
  
  • Initial 50,000 concept model: SENSUS/Ontosaurus
  
  • Knowledge acquisition & end-user modification:
  EXPECT




Problem Solving:





Natural
Language: 





Model Based Agents (Associate Systems):



    
  
  • Access to multiple, heterogeneous databases: SIMS 
  
  • Medical Critical Care Advisor: CritCare
  
  • Information Mediator for Webdata: ARIADNE 




Simulation:



    
  
  • Intelligent Forces: SOAR
  
  • Simulation Analysis and Explanation: PROBES




Multiagent Systems:



    
  
  • Teamwork-based cooperation among heterogeneous agents: Teamcore
  
  • Negotiation through argumentation: Dynamite




Education and Training:





Data Mining:DataCrystal



Simulation: RoboCup Simulation


Robots: YODA,
Robocup Robot Team, Conro






Automated Distance Education
(ADE)



The ADE effort is developing software that will allow students to take
courses at remote sites, either at home or at work sites. It uses a mixture
of World Wide Web and artificial intelligence technology to provide courseware
that responds dynamically to instructor commands and student needs. At present
the effort is working in two application areas: manufacturing education
and aircraft safety.



The ADE systems are built on a network of servers, some located at a
central school site and some located on individual student machines. Each
student machine is provided with a suite of browsing, communication, and
browsing tools, based on existing tools such as Hot Java and CU-See-Me and
multimedia authoring tools such as Director. Each query from the browser
is intercepted by the local server, which determines whether the requested
resource must be retrieved from the school server, or is available on the
local CD-ROM or hard disk. High-volume resources such as MPEG video can
be stored locally on the CD-ROM, eliminating the need to transmit such assets
over low-bandwidth telephone lines. Student interactions are automatically
recorded and uploaded to the central server, where they can be recorded
and analyzed.



The primary application of artificial intelligence in this framework
is in automated student assessment and curriculum sequencing. Students are
presented with exercises, which are graded automatically by the ADE software.
Using a Bayesian network approach, ADE computes the probability that various
course topics have been mastered, and generates an estimate for each course
module of the degree of benefit that the student will derive from taking
it. Course modules can then be selected and sequenced dynamically, in order
to meet the needs of individual students.






ARIADNE



Ariadne is a information mediator that extracts and integrates data
from semi-structured web sources.  Ariadne enables developers to
rapidly create ``information agents'' for the Web and to maintain
these agents as information sources change and new sources become
available.  Using Ariadne's modeling tools, an application developer
starts with a set of web sources -- semi-structured HTML pages, which
may be located at multiple web sites -- and creates a unified view of
these sources. Once the modeling process is complete, end users can
issue database-like queries as if the information were stored in a
single large database. Ariadne's query planner decomposes these
queries into a series of simpler queries, each of which can be
answered using a single HTML page, and then combines the responses to
create an answer to the original query.






DataCrystal



DataCrystal is a project that is applying advanced Data Mining technologies
to discover useful patterns and trends from very large data sets. It uses
a general mechanism called "metapatterns" to integrate deductive
database techniques, inductive data analysis tools, and human intuition
into a continuous discovery loop that is both interactive and autonomous.
It has been applied successfully in several real-world applications, including
discovering common-sense regularities from a large knowledge base, finding
circuit patterns from a telecommunication database, building prediction
models from a chemical research database, and constructing fault detection
rules from a semiconductor manufacturing control database.






Electronic Learning


The Electronic Learning project is developing tools to develop and deliver
advanced on-line higher education courses.  EL
is being conducted at the
Center for 
Advanced Research in Technology for Education (CARTE) at USC / ISI.



EL assists faculty in developing on-line courses.  In the process it
develops and tests new technologies for course management and
delivery, and reusable courseware components.  The results of evaluating these technologies are used to
drive further research.



Our areas of interest include:

    

  • Automated support for design and authoring of instructional materials

  • Automated assessment of student performance

  • Support or nonlinear curriculum models



We are currently engaged in a project to help put the introductory
circuits course in the Electrical Engineering Department on line.







EXPECT



The knowledge acquisition bottleneck is frequently cited as a major impediment
to broad dissemination of AI technology. The EXPECT project is addressing
that problem by developing a knowledge acquisition framework that empowers
users and domain experts to augment, modify and adapt knowledge based systems
without needing to understand the details of the system's implementation.
The key to EXPECT's approach is that it captures the design rationale for
knowledge based systems, and uses that design knowledge to guide a user
in augmenting the system. EXPECT has been used to build tools for transportation
planning and for air campaign planning.






Intelligent Documentation



This project is developing a documentation authoring tool called I-DOC
that automates the process of generating documentation and user help for
software systems. This tool will result in dramatic improvements in the
way documentation is developed, maintained, and used. Documentation is generated
dynamically, in response to specific requests for user information. When
the user requests information, the documentation system determines what
information content should be presented. It composes a response by combining
textual descriptions previously entered in a database, and automatically
generating natural language output to fill in the rest. The content of the
generated output depends upon user's intended use for the information, and
on the user's areas of expertise. This reduces the need for users to search
through volumes of irrelevant documentation in order to obtain answers to
specific questions. Interaction with the tool is much more akin to a question-answer
dialog than an access to a reference document. The work is based upon empirical
studies of question-answer dialogs between people about software systems.
Maintainers can update the documentary information as they viewing it, helping
to ensure that the documentation stays up to date.






Loom Knowledge
Representation Language



Loom is a language and environment for constructing intelligent applications.
The heart of Loom is a knowledge representation system that is used to provide
deductive support for the declarative portion of the Loom language. Declarative
knowledge in Loom consists of definitions, rules, facts, and default rules.
A deductive engine called a classifier utilizes forward-chaining, semantic
unification and object-oriented truth maintenance technologies in order
to compile the declarative knowledge into a network designed to efficiently
support on-line deductive query processing.



The Loom system implements a logic-based pattern matcher that drives
a production rule facility and a pattern-directed method dispatching facility
that supports the definition of object-oriented methods. The high degree
of integration between Loom's declarative and procedural components permits
programmers to utilize logic programming, production rule, and object-oriented
programming paradigms in a single application. Loom can also be used as
a deductive layer that overlays an ordinary CLOS network. In this mode,
users can obtain many of the benefits of using Loom without impacting the
function or performance of their CLOS-based applications.






MediaDoc


Complex software is difficult to modify in part because it is
difficult to understand; the greater the complexity the greater the
understanding problem.  MediaDoc is addresses this through the
development of 
software explanation
technology, i.e., the dynamic generation of presentations specific to
individual users' tasks and needs.  Software explanations reduce the
amount of time spent sifting through irrelevant information, helping
to ensure that the presented information is more readily understood.
MediaDoc presentations
integrate diagrams, animations, and
text.



MediaDoc offers several significant innovations over
existing graphically oriented software tools (e.g., CASE tools and
visual languages).  First, the generated presentations are
designed to meet specific communication objectives, highlighting
interesting system features and suppressing
repetitive details.  They explicitly employ graphical presentation
idioms used by expert graphical designers to convey information
effectively.  They seek to maintain a consistent level of complexity,
regardless of the complexity of the software being explained,
by limiting the amount of detail that is presented at any one time.
We are also exploring ways of making better use of spatial
metaphors to help users to orient themselves better when viewing
abstract diagrammatic presentations of software.
Media-Doc presentations in contrast will have tightly
integrated text, diagrams, and animation.  The presentations will be
interactive, enabling users to obtain additional information and
clarifications of the information presented.



MediaDoc is also concerned with extracting information from legacy
documents, so that it can be easily reused in dynamically generated presentations.





Parent
  Ed



The Parent Education Project at the 

Center for Advanced Research in
Technology for Education (CARTE) will be a collaboration with a number
of other institutions to develop electronic course materials that help
parents of pediatric cancer patients to deal with problems that they encounter.








Penman/Pangloss



The Natural Language project at USC/ISI has been conducting research
in various areas of Computational Linguistics / Natural Language Processing
since 1978. This research includes work on single-sentence realization,
multi-sentence text and sentence planning for descriptions and explanations,
parsing and semantic analysis, the semi-automated construction of large
semantic knowledge bases (so-called ontologies) and lexicons of various
languages, and the automated planning of multimedia human-computer communication.



Since 1991, much of this work has been used in the construction of Machine
Translation systems. One of these, JAPANGLOSS, translates Japanese newspaper
articles in any domain into English using a mixture of symbolic and statistical
techniques. Statistical techniques provide large-scale coverage at a lower
level of quality while symbolic (linguistic and other traditional) techniques
provide reduced coverage of the language but at higher quality. JAPANGLOSS
is used in both an e-mail translation server and a prototype translating
copy machine, the latter incorporating optical character recognition. Another
MT system is SPANGLOSS, built in collaboration with research groups at Carnegie
Mellon University and New Mexico State University, that translates Spanish
newspaper texts into English.



Over a decade of previous work on sentence generation resulted in the
PENMAN English sentence generator, through which the project is linked to
sister projects in Germany and Australia. PENMAN (or its multilingual descendant
KPML) is being used in several projects in North America, Europe, and Asia.






PROBES



PROBES is developing analysis and explanation capabilities for training
simulations, in order to make such simulations more effective as training
tools. It is part of the Exercise Management effort within DARPA's Computer-Aided
Education and Training Initiative.



Military training simulations typically incorporate large numbers of
entities, both human and computer-generated, which interact in a simulated
world. Setting up such simulations is quite laborious, and once the simulation
is running it must be monitored closely in order to assess whether the simulation
is providing the human participants with the intended training experiences,
and whether these experiences are having the desired effect. PROBES brings
plan recognition and explanation technology to bear on this problem. Assuming
that the training and mission objectives have been specified as part of
the exercise scenario, PROBES observes entity behavior and recognizes behavior
patterns (plans) that are relevant to those training objectives. It also
provides dialog interface to entities, enabling trainees to interact with
the simulated scenario participants in "after action reviews."
Trainees can find out from the participants why they did what they did.
Integrating the plan recognition and explanation capabilities yields "talking
probes," i.e., intelligent agents that can monitor activity in the
simulation and provide the trainee with narrations. and critiques.



It is expected that the project will extend ideas developed in the Debriefable Agents
project, and coordinate closely with work in the VET
project.






SIMS



The overall goal of the SIMS project is to provide intelligent access
to heterogeneous, distributed information sources (databases, knowledge
bases, flat files, programs, etc.), while insulating human users and application
programs from the need to be aware of the location of the sources, their
query languages, organization, size, etc.



The standard approach to this problem has been to construct a global
schema that relates all the information in the different sources and to
have the user pose queries against this global schema or various views of
it. The problem with this approach is that integrating the schemas is typically
very difficult, and any changes to existing data sources or the addition
of new ones requires a substantial, if not complete, repetition of the schema
integration process. In addition, this standard approach is not suitable
for including information sources that are not databases.



SIMS provides an alternative approach. A model of the application domain
is created, using a knowledge representation system to establish a fixed
vocabulary describing objects in the domain, their attributes and relationships
among them. SIMS accepts queries in this high-level uniform language. It
processes these queries in a manner hidden from the user, ultimately returning
the requested data. Thus, the queries to SIMS need not contain information
describing which sources are relevant to finding their answers or where
they are located. Queries do not need to state how information obtained
from different sources should be joined or otherwise combined or manipulated.
SIMS uses a planning system to determine how to efficiently and transparently
retrieve and integrate the data necessary to answer a query.






Soar



The Soar project is an interdisciplinary, multi-site attempt at
constructing autonomous intelligent agents capable of exhibiting effective
behavior across a wide range of tasks, from the highly routine to the
complex and open-ended.  This involves investigating a variety of
individual intelligent capabilities as well as the integration of such
capabilities into a unified architecture and (groups of) implemented
agents.  The Soar architecture already embodies a number of the
capabilities required of individual intelligent agents, such as reactive
behavior, general problem solving, knowledge-based decision making,
reflection, and learning.  We continue to investigate such capabilities,
and to maintain a keen interest in key technological issues such as
efficient production match.  However, recent experiments with Soar-based
intelligent agents in virtual environments revealed Soar to be particularly
deficient in a number of the social capabilities required to perform
effectively in multi-agent environments.  So, our primary research focus
over the past few years has been on social capabilities, such as the
multi-agent aspects of perception, understanding, learning, planning, and
(coordinated) execution.  Over the next few years we expect to continue
working on these topics, as well as to also start looking at other aspects
of social interaction, such as emotional interactions.

The principal domain within which we are implementing Soar-based
intelligent agents is the synthetic battlefield -- a distributed interactive
simulation environment intended for analysis, development, training and
rehearsal purposes -- where we are developing pilots and commanders for a
range of helicopters and missions (after earlier working on fixed-wing
aircraft pilots).  However, we are also developing a team of synthetic
soccer players (which placed 3rd in a field of 32 at the RoboCup '97
tournament) , and the VET project is developing Soar-based tutors for
virtual environments.  Additional application domains may also be in the
offing.






STEVECO


The STEVECO project is a collaboration between USC's

CARTE and Intelligent Systems Technology, Inc.  Its purpose is to
develop productizable pedagogical agent
technology.  
There is close collaboration with the
VET project, which
has been conducting research in pedagogical agents, and the
ADE project, which
is also developing commercial grade agent technology for distance
learning.





Trauma Care Information Management
System (TCIMS)



Early and aggressive combat casualty or trauma care offers the potential
to reduce deaths and minimize crippling disabilities. The Advanced Research
Projects Agency is sponsoring a Trauma Care Information Management System
(TCIMS) to enhance the decision making abilities of Military Battlefield
Medicine personnel and to facilitate information flow and patient processing
in battlefield medicine and civilian trauma care situations. It is designed
to share critical trauma care information between military and civilian
responders during joint relief efforts for national disasters, such as the
recent California earthquake or Florida hurricane.






Virtual Environments for
Training (VET)



The primary objective of the VET project is to try to answer the following
question: How can virtual environment technology be used most effectively
for training? It is clear that virtual reality has significant potential
for education and training, since it allows students to experience situations
which are costly or dangerous to recreate in real life. However, there are
many unanswered questions regarding the design of virtual reality training
systems. How should a virtual environment be simplified or enhanced in order
best to facilitate learning? And how can AI technologies such as autonomous
agents, plan recognition, or explanation increase the effectiveness of a
VR training system?



ISI has joined a consortium including Lockheed Martin AI Center and USC
Behavioral Technologies Laboratory in order to investigate these questions.
As part of this effort ISI will be investigating the role of AI technologies
in virtual environments. In particular, we will be investigating the effectiveness
of pedagogical agents in fostering learning. Pedagogical agents are
autonomous agents that take the form of animated characters, and which can
interact with human trainees within the virtual environment. Unlike the
conventional coach in an intelligent tutoring system, which typically plays
the role of critic or kibitzer, pedagogical agents can involve themselves
more directly in the activity, as a partner in the problem solving process.
Pedagogical agents should be able to show trainees how to solve problems,
recognize when they are experiencing difficulties, and collaborate in completing
the task. We anticipate that by populating virtual environments with pedagogical
agents, we can make environments more effective as learning tools, as well
as more engaging experiences. We also conjecture that the activity of developing
cognitive models of skill will in turn determine what elements of the environment
are involved in completing the task, and hence help determine the degree
of fidelity in which the real world must be modeled in the VE.







[ Division Overview | Projects
| Publications | People
| ISI Home ] 

[ Employment
Opportunitites | AI Seminar | AI Grads ]


[  The InSiDer
| Div 3 Admin |  ISD Helpdesk  ]