CS 8113D


Notes by D. Jones
October 25, 1994


PHI, A Logic-Based Tool for Intelligent Help Systems
	Bauer, Biundo, Dengler, Koehler, Paul

Intelligent User Support in Graphical User Interfaces
	Fehrle, Thies, Berger

---
The PHI paper introduced varying levels of help:
	Passive help
	Active help
	Cooperative help
	Implicit help.

There are two types of planning, first principle planning and second
principle planning.  With first principle planning, you use a plan
language to start from scratch and generate a sequence of steps.
Second principle planning involves reusing solutions to old problems.

The papers mentioned the use of artificial intelligence techniques,
but went into very little technical detail -- especially the PLUS
paper.  The PHI paper used Prolog, which centers on theorem proving
based on axioms.  One difference provided by the AI approach is the
use of temporal logic, where logical values have a notion of time and
duration.  A value may become true at a certain time, and may again
change after another period of time.

One problem noticed with the papers is that the target environments
seem like "toy" systems, and there is a question of scalability.  For
example, in the PHI paper they are implementing an email program.  The
ideas are also somewhat outdated, as modern AI techniques have reuse
as a primary goal.  The algorithms are also less opportunistic, and do
not take advantage of situations as they come up.  They use the closed
world assumption: what you don't know (or can't prove) is assumed
false.

How far can you go with the user interface issue?  You would need a
really sophisticated user model to get anything useful, and there is a
lot of extra complication introduced by the AI.

The key to the PHI paper is the enumeration process whereby you
enumerate plans.  Logic planning is used by the PLUS paper.  IBM has
already patented a variation of PLUS, and is using a user interface
management system (UIMS) designed at IBM.

Another problem raised is that you can't just model WHAT is going on
without WHY it is going on.  However, it is possible to find out the
final goal of the user.  It is very difficult to enumerate all the
plans because there is a combinatorial explosion.  It may be possible
to store all the enumerated plans, but the search time would be n(log
n).  You want to anticipate what the user wants on the spot, but this
impossible in larger systems.

How far can you optimize the system?  At what point does the added
cost exceed the benefits?  (an extra inch could cost a mile) The
designer's intent of the system may or may not overlap with the user's
use of the system.

A simpler approach to anticipating user's actions is to use
statistical methods, as opposed to an actual "understanding" by the
computer of the users intentions.  The user usually doesn't care if
the system "understands" their desires -- only that it does what is
best for him/her.  If the actions of the system disturb the work of
the user, then it is bad.  However, if the user requests help or hints
about what to do next, and the computer makes suggestions based on
incomplete information, this is not bad because it gives the user
direction.