These are the ASCII versions of the slides from my presentation on
Friday, Oct 7 on Hermes and Schlimmer.

In addition, there were four more that were copied from the paper 
itself.

1. Contextual Mode for Powerbooks
2. Button Box mode for Patterns
3. FSM for Patterns
4. Rules for merging token list into FSM


Thanks. 

wjones@cc


===========================================================


Software Agents:
Completing Patterns and Constructing User Interfaces

-----------------------------------------------------------

The Note-Taking System
    o Paper Paradigm in a Digital World
    o Pen-based systems mark a shift toward paper paradigm
    o Note-taking system designed to speed up and reduce
      errors in data entry by
       -  Predicting what the user is going to write
       -  Construct a graphical user interface on request

-----------------------------------------------------------

Software Characterization
    o Specificity of task
       -  Task Independent > Generic > Specialized
    o User Customization Required
       -  None > Generic > Complete Customization
    o Self-Customizing Software tends to move toward low user
      customization and generic specificity

-----------------------------------------------------------

Definitions
    o Note
       -  A short sequence of descriptive terms that describe an object of 
interest.
    o Agent
       -  Software that allows the user to enter data in free-form and
          organizes the info behind the scenes.

-----------------------------------------------------------

Modus Operendi
    o Two Modes
       -  Contextual        -  Free Form
       -  Interactive GUI   -  Button Box
    o Completion Button
    o Mode Switch
    o Color Coded Hints
       -  Green  = Confident
       -  White  = Duh

-----------------------------------------------------------

How it works
    o Data entered in contextual mode parsed into list of tokens.
    o Tokens merged into a Finite State Machine (FSM)
    o GUI constructed from FSM
    o Predictions come from decision tree classifiers at states
      within the FSM

-----------------------------------------------------------

Learning the FSM
    o Heuristic to merge token list into FSM
    o Multiple domains

-----------------------------------------------------------

Learning Classifiers and Prompting
    o A decision tree developed for each state
    o Prediction based on frequency of choice, total visitations,
      completeness of token list
    o From empirical study, prediction is best when there are a
      lot of constant and repeating strings
    o Other mechanisms may provide better results

-----------------------------------------------------------

Constructing the GUI
    o Available when each state is mapped twice
    o Any path that loops back to its original state is a check box
    o Non-looping paths become a set of radio buttons
    o Labels built by concatenating tokens along path
    o Some paths can't be rendered before other choices are made

-----------------------------------------------------------

Observations/Limitations
    o The system is tabula rosa
    o Domain-specific knowledge would make it more useful in 
      the real world
    o Prototype not built on target machine class
    o Generalized strings of numbers or chars
    o FSM forces syntax, constrains order of interface
    o Designers winged some parts

-----------------------------------------------------------

Criteria for Evaluating Agents
    o Does system anticipate the user's actions?
    o Is it GUI or CLI?
    o Can user override the system?
    o If there are multiple modes, does the system force one
      over the other?
    o Does the system learn as it goes?
    o Can the user fine tune the system?

-----------------------------------------------------------

Discussion Topics
    o Is this a useful system?
    o Does the user train the system or vice versa?
    o Is always adhering to the paper paradigm a good thing?

-- 
					
			-William Raymond Jones II (wjones@cc.gatech.edu)