perm filename CACHE.OLD[AM,DBL] blob
sn#392777 filedate 1978-11-07 generic text, type T, neo UTF8
6-Nov-78 10:44:48-PST,6201;000000000001
Mail from RAND-UNIX rcvd at 6-Nov-78 1044-PST
From: Rick at Rand-Unix
Date: 6 Nov 1978 at 1046-PST
Message-Id: <[Rand-Unix] 6-Nov-78 10:46:25.rick>
To: lenat @ sumex-aim, klahr
cc: rick
Subject:Joint IJCAI paper
Phil and Doug:
I've given lots of thought to what we are setting out to do
on this paper. And the more I think on't the greater get my aspirations.
What do you think of the following proposed joint venture?
Title: Cognitive Economics
[or Principles of Cognitive Economics]
[or Cognitive Economy Revisited]
Thesis: Intelligent systems explore very small subsets of their potential
external and conceptual worlds. They must develop efficient forms
of representation and operation to increase their capacities.
Some of these forms involve abstraction, caching, and expectation-
simplified processing. These capabilities in turn can combine to
provide extremely powerful increases in performance. For example,
all three can combine to simplify simulation or, one of its
related functions, detection of surprising events. Our
analysis of the economic principles of modern AI systems or
(presumably more sophisticated) human intelligence suggests that
previous ideas regarding cognitive efficiency have erred in
fundamental ways. For example, the nonredundant storage of
properties in hierarchical inheritance nets increases many processing
costs while providing minimal storage cost savings. We propose
methods to exploit the potential advantages of both schemes.
Outline:
1. Introduction
Our model of intelligent system organization
Concepts, heuristics, PDIS
Our model of intelligence: knowledge and its expansion (through
experimentation, discovery, conjecture, conditioning)
Our model of computing:
Cheap storage, expensive knowledge acquisition, limited
computing cycles
The problem:
Want to develop initial systems quickly and then have
them speed up and economize their computing to maximize
their potential
The principal ideas:
Abstraction
Caching
Expectation-simplified computing
e.g., ignoring expcted data
giving priority to surprising data
feeding back to confirmed/disconfirmed predictions
Outline of the rest of the paper
2. Abstraction
Desirability of being able to compute rough answers cheaply
Conceptual hierarchies
Heuristic Hierarchies
Interpretation and planning at levels of abstraction
Eg., rules of bomber simulation at difft levels
(this example will ultimately be used to suggest
caching for simplification)
Related research
3. Caching
Modifying memory to save computed results to speed subsequent accesses
Generalization of hardware concept
EURISKO types of caching, as first examples
Contrast with psychological conjectures of cognitive economy
(e.g., Collins&Quillian, KRL, ...). More like HR↑2 Plasticity
model of storing all retrieved paths as direct links
General principles
Updating Principles
-------------------
When
Why
How
get demon traps that flag the cache as out of date
the user requests updating if the cache seems staleness
Where
In what form
What
When not to
How to
Storage Principles
------------------
When
Every time you have to call a lower order function to eval. it
& it took quite a while.
You've caled it before, recently & the value didn't change.
Why
Cost of recomputing vs. cost of storage.
Context of subsequent cals is similar enough (e.g.l, the same
arguments will come up again.
How
Called functions might suggest how to cache their value in higher
calling caches (e.g., my value changes often so cache my defn.).
Cache should be transparent & discardable (should be able to throw
them all away if space needed).
Where
In what form
value ) what level of abstraction (partially evaluated
expression) symbolic expression)
Stack previous values to enable you to tell if they're changing.
What
You store a flag saying you've been here before.
When it was computed.
How much effort was expended on it, down to what levels of
algorithms, with what around caches incorporated.
Certainty of the result.
When not to
The value changes too frequently.
The function evaluates as fast as the caching mechanism itself
Space is too tight
How to eliminate caches
Space tight--> eliminate last used caches (last referenced)
4. Expectation
Central notion: reserve your computing for opportunities to realize
potential for expanding knowledge
You may decide how much to expend re-confirming the expected
Reductions realizable through expectations:
Perceptual set: see what you expect with less effort
Surprise: heighten sensitivity to data inconsistent with
expectations
Predict and prepare
What mechanisms are implicated?
Caching
PDMs (as triggers or demons)
Relevance to learning
Confirm or disconfirm predictors
This requires setting up PDMs to fire on dis/confirmation
5. Cognitive economy revisited
Sample problem: using a world model (simulator) to answer questions
(e.g., what'd happen if 100 bombers went in there?)
Representation of this knowledge as PDMs at difft levels of abstn
Ability to generalize and cache results at one level at the
next higher level,
e.g. either as numerical tables, stat. distns, or
symbolic expressions
Ability to answer some questions appropriately for the requestor
at a high level of abstraction
KB Design
One good reason to use inheritanc is to speed knowledge
implementation, not computing performance
Using the system should result in its speedup
Storage should be cheap
Machine architecture
PDI should be cheap
PDMs should be scheduled with variable resources and
should be able to spend effort accordingly
How could propagation of changes be made efficient?
6-Nov-78 10:55:54-PST,1738;000000000001
Mail from RAND-UNIX rcvd at 6-Nov-78 1055-PST
From: Rick at Rand-Unix
Date: 6 Nov 1978 at 1057-PST
Message-Id: <[Rand-Unix] 6-Nov-78 10:57:41.rick>
To: lenat@aim
Subject: EPA Interests in rules for predicting chemical pathogens
Doug-- I inadvertantly omitted you from this msg list. -- Rick
------- Forwarded Message
From: Rick at Rand-Unix
Date: 6 Nov 1978 at 1053-PST
Message-Id: <[Rand-Unix] 6-Nov-78 10:53:59.rick>
To: Feigenbaum@sumex-aim
cc: Fagan@sumex-aim, Rha, Gaines, Rick
Subject: EPA Interests in rules for predicting chemical pathogens
Ed-
Both as director of HPP and Rand's preeminent AIM consultant,
you should probably look into EPA's interest in being able to
conjecture likely dangers among the 200 or so new organic chemicals
that are produced daily. Their basic problem is that they don't have
the foggiest idea how to discern dangerous new chemical compounds and
they are literally swamped with new creations. My idea is to combine
the kinds of learning methods developed in meta-dendral and by Lenat and
myself to conjecture rules of constituency, structure, etc.
Considering that NIH and EPA are holding a two-day preliminary workshop
on this and are thinking only of cluster analysis, pattern recognition
and similar techniques, it seems an idea opportunity for such AI methods.
I have learned about this only through the grapevine. The right person
to talk to I learned is Steve Heller 202-755-0881 at EPA. If you
agree that this looks promising, would you give him a call? I'm
interested in this but it seems likely that your interests are more
central and strongly connected in this area.
Best wishes,
Rick
------- End of Forwarded Message
7-Nov-78 16:26:56-PST,2430;000000000000
Mail from RAND-UNIX rcvd at 7-Nov-78 1626-PST
From: Klahr at Rand-Unix
Date: 7 Nov 1978 at 1628-PST
Message-Id: <[Rand-Unix] 7-Nov-78 16:28:14.klahr>
To: rick
cc: lenat @ sumex-aim
Subject: Joint IJCAI paper
Rick,
Based on an initial scan of your outline, I think it looks
great. Some preliminary thoughts:
1. Title: Cognitive Economy in Artificial Intelligent Systems
2. Abstraction: I don't think we should delve into the military
domain of bomber simulations for the IJCAI paper. It
may turn off alot of people. The hearts domain may
similarly turn off a different group. However, since
we'll have examples from EURISKO, examples from hearts
should provide an additional example domain and will
not look central to the ideas presented. The area of
abstraction in simulations is powerful as we are, and
will, experience. This is perhaps worthy of its
own paper. If we do want to talk about abstraction
in simulations, I suggest alternative domains, eg,
ship or air traffic control, sporting events (football
strategies), international terrorism (more people
sympathetic here), appointment scheduling (as in proposal),
etc.
3. Caching: looks good.
4. Expectation-simplified processing: This is a confusing phrase.
The only alternative I can think of now is expectation-
focusing. The economy here is in terms of subsequent
analysis of good/bad consequences. Expectation-focusing
directs the analysis and diagnosis of behavior to those
heuristics that were instrumental in the resulting behavior.
Demons can be generated by heuristics to fire when the
heuristics' expectations are met/rejected. Thus demons
economize here by pinpointing heuristics that impacted
the resulting behavior.
5. The idea of forming and storing symbolic expressions deserves to
be a fourth independent cognitive economy. I feel it
is much more than a simple cache. Forming these expressions
may involve considerable pattern-matching, deductions, etc.
The resulting expression incorporates more than a simple
evaluation (such as caching functional values). In fact,
I would consider it to involve learning as well as caching.
The expressions are really meta-evaluations.
I think the paper would be a significant contribution to IJCAI. I will
be happy to participate in it.
--Phil