RE: HM.applications-Profiling-Level of Details/Abstraction

["Bullard, Claude L (Len)" <clbullar@ingr.com>]

Yes.  I would certainly like to see more input from others 
but phase 0 with 70 lurkers and six active participants 
proves to me that most folks are easy riding and the 
others are playing poker.  Life is short.

Theories such as gestalt are interpretive domains.  In 
effect, they name patterns and the system uses these patterns 
to select some attributes of the subject domain (the situation, 
communication, etc. being interpreted).  HumanML high level 
classes are used to group these interpretive domains to enable 
alternative interpretations to be selected.  It is useful to 
know that communication and learning are tightly intertwined.

HumanML isn't precisely about human-to-human 
communications.  It is also a digital means to enhance such 
communications through translation, problem solving, 
selecting representations, and enabling representations 
eg, given a HumanML knowledge base, an avatar can 
direct an HCI interaction or a system can create a 
context-appropriate interface.  So we have to consider 
HCI problems that HumanML can help solve (consider 
usability or learning): human to computer to human.

Working with the Amodeus project papers, here is an example 
of a learning based interaction that points out some of the 
issues of a learning process of HCI.  In this example, I am 
only looking at the human-to-machine given that one will learn 
to use human to machine tools first before working the other 
side of that communication.
 
Much reasoning given a situation of uncertainty is analogical.
Analogy is inductive and needs a verification action has desired 
effect to establish a rule.  Knowledge in the HumanML model is 
contained in declarative representation and in rules.  To 
use induction, the process is:

1. Process sets goal 
2. Looks for rule.  If no rule found
3. Identify similar situation using surface similarity of 
   goal situation and experience. (source and target)
4. Identify an action from situation. 
5. Test action.
6. If action is successful, create new rule 

-  If object is this.object and goal is this.goal; use this.action.

This requires the system (eg, the Human object) to identify 
the meaningful attributes.

SELECT objects WHERE object is-a objectType AND has-name (property =
somevalue) 

The facts may be episodic (was started) instead of categorical 
(is-a object).

It may need chains of facts between the example task and the example action,

to make features of the action meaningful.  What identifies the correct 
action/control in a set (the item selection problem)?  

Scenario:  a human unfamiliar with a car's controls wants to 
cool the car.

Goal is "cold"

Given

Episodic fact:  (this.action(push red button) -> this.state(heater.On) ->
Result = "hot" )
Context fact:   (blue = "cold")

Learns:

- rules for learning 
  (this.action (this.object(controls) -> this.state(system.on))
- rules for specific results 
  (this.action (push blue button) -> this.state (airConditioner.on) ->
Result="cold")

and these become part of the knowledge base.

Sean, how would RDF represent this?  Then we should inquire as to how a 
human object uses the knowledge base.

Len 
http://www.mp3.com/LenBullard

Ekam sat.h, Vipraah bahudhaa vadanti.
Daamyata. Datta. Dayadhvam.h


-----Original Message-----
From: Rex Brooks [mailto:rexb@starbourne.com]

I'm actually not snipping this because this is essentially what I was 
getting at, although I would like to hear from a few more voices on 
their takes... As we move from the general to the specific with 
profiling, we need to reach consensus on the most useful common 
general categories. I'd just like to collect some enumerations for 
our core documents, as a place to start.