Professional Documents
Culture Documents
1
VRAI Group – EPFL – 1015 Lausanne – Switzerland
2
AIC – SRI International – Menlo Park, CA – USA
Today computers are part of the Active based assistant for endoscopic
standard equipment of modern surgery neurosurgery is under development.
rooms. They assist surgeons in Using natural modalities such as speech
performing complex procedures that recognition and hand gestures, it enables
would not be possible otherwise. surgeons to interact with computer
However, despite the availability of based equipments of the operating room
more powerful and complex computer as if they were full active members of
systems, their user interfaces have not the team. In a broader context, Active
been adapted to fully leverage their aims to ease the development of
potential. A new type of software, intelligent software by making required
behaving as an independent intelligent technologies more accessible. It will
assistant, is needed to better assist help foster research innovation, easier
surgeons and their staff. Building an development cycle and deployment of
intelligent assistant is a difficult task that this new type of applications.
requires expertise in many fields ranging
from artificial intelligence to core
software and hardware engineering. We
INTRODUCTION
Although computer systems have grown
believe that providing a unified tool and
in power, access more networked
methodology to create intelligent
content and services, computer
software will bring many benefits to this
interfaces have not changed.
area of research. Our solution, the
Conventional user interfaces with simple
Active framework, introduces the
direct manipulation commands are no
original concept of Active Ontologies to
longer sufficient to fully leverage such
model and implement intelligent
rich and dynamic environment [1]. The
applications. Based on suggestions and
medical field is no exception.
constant evaluations from surgeons, an
Figure 1 : Active Editor
Computers are now part of the standard Reaction, decision making strategies and
equipment used in modern surgery complex task execution are the
rooms. To fully leverage this new responsibility of planning systems.
context, modern software systems Finally, as planning unfolds various
should behave as intelligent assistants actions are taken by the system. Based
able to observe and sense their on their nature and purpose, intelligent
environment, for instance human inputs, systems act through a wide range of
to analyze a situation by mapping input modalities. They communicate with
senses into a model of what tasks and humans, gather information or
events may be happening [2]. They physically change their environment.
would then understand and anticipate Designing and implementing intelligent
what the user might need to finally act assistants software is also a difficult
to produce relevant and useful task.
behaviour. The development of Due to the variety and complexity of
intelligent assistants requires expertise technologies required, intelligent
in many ¯fields [3]. assistants are made of a collection of
Perception of human activities is components written in many different
typically based on techniques such as programming languages. Connecting
computer vision or speech recognition. various heterogeneous programs,
Understanding the meaning of input sometimes remotely, requires strong
signals, is performed by natural technical knowledge and careful
language processors, dialog systems or deployment policies. Testing and
activity recognition mechanisms. debugging distributed heterogeneous
systems is also a complex task. To information and services to help users
identify and correct bugs, events and with complex tasks [5]. Scheduling
associated values need to be tracked meetings, managing an agenda and
from one component to another. Finally, communicating also represent
combining many different approaches, applications where intelligent assistants
tools and technologies limits the overall are relevant [6].
performance and extensibility of the Intelligent assistant are also relevant in
system. the domain of heterogeneous smart
We believe that providing a unified tool spaces, instrumented rooms able to
and methodology to create intelligent sense their environment and act upon
software will solve many of the events and conditions. In the surgical
problems described above and bring field, modern operating rooms are
many benefits to this area of research. It becoming such smart spaces. Many
will allow more researchers and components can now be connected and
engineers to work in the field by controlled so that intelligent assistant
providing a bridge between core AI software can be deployed to assist
technologies and practical engineering. surgeons and their staff. Existing smart
This paper introduces our spaces projects are designed and
implementation of this vision, the optimized for specific domains,
Active framework. The next section is implemented using proprietary
dedicated to related work on building frameworks and methods. Our goal is to
intelligent assistants. The section Active provide a more generic intelligent
Framework outlines the Active original system toolkit, composed of a suite of
concepts, architecture and current tools and methodologies to rapidly
implementation. The next section design and deploy complex software
presents how the Active framework is into smart spaces.
used to implement an intelligent Our work also relates to the field of
assistant in the context of neurosurgery. multi agent framework research. In this
Finally, a conclusion presents directions area, heterogeneous existing AI based
of our future work. components are turned into agents able
to form communities working together
RELATED WORK with humans to help them solve
problems. In this context, the open agent
architecture [7] OAA introduces the
By definition, intelligent interactive
powerful concept of delegated
systems are based on various AI
computing. Requests and plans are
techniques.
delegated to a facilitator in charge of
Relevant efforts related to our research
orchestrating actions based on declared
can be classified into three categories.
capabilities of agents. Thanks to its ease
First, the area of interface agents aims at
of deployment and clean design, OAA is
creating intelligent user interfaces to
used in a large number of projects.
assist humans in specific domains [4].
Though very powerful, OAA does not
For instance, the Internet is an
provide a unified methodology to create
environment where intelligent assistants
intelligent systems. It rather provides a
can leverage a vast amount of
framework where heterogeneous [10] [11] are available and have proved
elements, written in many programming their relevance in the field of intelligent
languages, are turned into OAA systems. BDI based engines would be
compatible agents to form intelligent well suited to be the core of our
communities. Similarly, the Retsina [8] research, where dynamic decisions need
framework is advanced multi agent to be made to respond to an event. Their
architecture to build distributed design is nevertheless constrained to
intelligent systems. It is based on four dynamic planning and would not be
classes of agents. Interface agents that suited to implement tasks such as
interact with users, task agents that carry natural language processing or modality
out plans, information retrieval agents fusion.
and middle agents to help match agents
that request services with agents that
provide services. Though very efficient
ACTIVE FRAMEWORK
in producing independent reactive
behavior, Restina would not be suited as 1. Conceptual Overview
a unified methodology to implement Our solution, the Active framework,
basic AI components such as natural provides a unified tool and methodology
language processors or multimodal to eases the development of intelligent
fusion engines. In addition the design of software. Active is based on the original
Retsina uses different formalisms for concept of Active Ontologies, used to
communication, domain representation model and implement applications. A
and reasoning technique. In contrast, our conventional ontology is defined as a
aim is to use the same formalism for all formal representation for domain
intelligent assistant aspects. knowledge, with distinct classes,
Finally, undertaking tasks on behalf of a attributes, and relations among classes;
user and attempting to understand what it is a data structure. An Active
actions are being carried out involves Ontology is a processing formalism
planning. BDI based systems [9] where distinct processing elements are
provide goal oriented reactive planning arranged according to ontology notions;
in dynamic and partially known it is an execution environment. An
environments. Beliefs represent the Active Ontology is made up of
model and state of the world and a plan interconnected processing elements
library defines how to achieve goals. called Concepts, graphically arranged to
Intentions are activated plans elected represent the domain objects, events,
and picked from the library to reach actions, and processes that make up an
some goals. The list of intentions is application. Concepts communicate with
constantly evaluated with beliefs, thus each other through channels, passing
providing a reactive behavior to the state information, hypotheses, and
system. Many BDI implementations requests.
Figure 2 : Active Application Design
[1] MAES P., Sardine: dynamic seller [9] RAO A.S, GEORGEFF
Agents that reduce work and strategies in an auction M.P.
information overload marketplace BDI-agents: from theory to
Communications of the ACM, ACM Conference on Electronic practice
1995, 38. Commerce. 2000, 128-134. Proceedings of the First Intl.
Conference on Multiagent
[2] SOWA J.F., [6] BERRY P., MYERS K., Systems, San Francisco, 1995.
Architecures for intelligent URIBE T., YORKE-SMITH N. [10]MYERS K..
systems. Special Issue on Constraint solving experience A procedural knowledge
Arti¯cial Intelli- with the calo project approach to task-level control
gence of the IBM Systems Proceedings of CP05 Workshop In proceedings AIPS-96, 1996,
Journal, 2002, 41 : 331-349. on Constraint Solving under AAAI Press 1996 158-165
Change and Uncertainty, Sitges, [11] NORLING E., RITTER
[3] WINIKOFF M., Spain, 2005 4-8 F.E.
PADGHAM, L. HARLAND. Embodying the JACK agent
Simplifying the development of [7] CHEYER A., MARTIN D. architecture
intelligent agents The open agent architecture Australian Joint Conference on
Australian Joint Conference on Journal of Autonomous Agents Artificial Intelligence. 2001, 368-
Artificial Intelligence, 2001, 557- and Multi-Agent Systems. 2001, 377.
568. 4(1) : 143-148.
[12] GRAETZEL C., FONG
[4] MIDDLETON S.E. [8] SYCARA K., DECKER K., T.W, GRANGE S., BAUR, C.
Interface agents: A review of the PANNU A.S., A non-contact mouse for
field, 2002. WILLIAMSON,.M, ZENG D. surgeon-
Distributed intelligent agents computer interaction
[5] MORRIS J., REE P.,MAE IEEE Expert, 1996 Technology and Health Care
P. 2004, 12(3) : 245-257