As it was described in section 4, INSPECT interfaces
with the user through ACPT. It receives most of its inputs (such as
the plan to be inspected) from ACPT, and it returns its output to be
displayed by ACPT. Therefore, the two systems were built with a tight
integration. The schema for integration is shown on
Figure 2.
[IMAGE ] Figure 2: Schema of the integration between
INSPECT and ACPT.
The integration between INSPECT and ACPT makes use of the ACPT
architecture. Its main component is the ACPT Foreign Agent, a
program that handles all the communication. The ACPT Foreign Agent is
composed of two parts. The first is a NOM client (developed in C++) to
exchange data with ACPT by communicating with ACPT's NOM server. The
second is the CLOS repository controller, a CommonLisp program that
calls and is called by INSPECT, and whose main function is to build,
in the CommonLisp image, a mirror of the current plan and the current
INSPECT preferences. These data are both edited and stored in ACPT.
The CommonLisp mirror stores the information as CLOS objects, which
simplifies the translation from the C++ object format used by ACPT.
The communication between the NOM server and client is made via
TCP/IP. The communication between the NOM client and the CLOS
repository controller is made by linking the C++ program to the
CommonLisp image using the Lisp foreign function interface. The
connections between the Foreign Agent and INSPECT, as well as within
INSPECT, are all made within a single CommonLisp image.
Since INSPECT needs the information to be stored in the Loom knowledge
representation language, the second main component is a group of
CommonLisp modules that translate the objects in the CLOS mirror into
appropriate Loom structures (more specifically, they are represented
as Loom facts). Both the data about the plans and the preferences are
integrated into INSPECT's knowledge base.
An execution of INSPECT occurs basically as follows:
The user employs the appropriate ACPT interfaces for specifying
the preferences for INSPECT's execution. This step is optional;
default values are used if necessary.
After having built (part of) an air campaign plan, the user uses
the ACPT interface to request that the plan is checked by INSPECT. A
message is sent (through the NOM server) to the foreign agent to
execute this request.
The foreign agent exchanges the appropriate messages with the
NOM server and creates the necessary CLOS objects in order to build
the mirror of the current plan and the INSPECT preferences on the
Lisp side. When this translation is ready, the foreign agent
requests INSPECT to check the plan.
INSPECT translates the CLOS objects to Loom facts in its
knowledge base. The INSPECT reasoner then scans the plan for
problems, obeying the specified preferences. For each problem found,
an agenda item is created. These items are then ordered by type
(warning, caution, note), grouped into an agenda object and the set
is translated into CLOS objects which can be read by the ACPT-Lisp
interface. This agenda includes already the explanations and fixes
for each of the agenda items.
Control is regained by the foreign agent, which sends the
appropriate message to the NOM server so as to update the agenda,
which can be in turn displayed by ACPT and browsed by the ACPT user.
Without any plan or FOB loaded, the current KB has
aproximately 360 concepts, 120 relations and 190 instances. Also,
there are 35 EXPECT problem solving methods. There are several other
components as well, e.g. tables, definitions of explanations and
fixes, which are built in a different way and are difficult to
quantify.
performance
In order to give a rough idea of execution times,
the following are approximate data about the execution of the above
cycle for the scenario which was employed in IFD-4. The whole
execution takes about four and a half minutes, distributed roughly
as follows: two minutes for creating the CLOS mirror of the plan and
preferences, one minute to translate the CLOS objects into Loom, one
minute for INSPECT to scan the plan and generate the agenda, half
minute to send the agenda back to ACPT.
knowledge engineering and acquisition
The knowledge engineering
effort in building the IFD-4 version of INSPECT was considerable. We
had about 50 hours of interviews with Checkmate and ISX experts, and
about the same amount of time preparing for the interviews. It took
on the order of 250 person-hours of knowledge engineering, not
including system building.
![[IMAGE ]](img2.gif){kind=link}