PIP-Bullet2-Cognitive Information Processing Technology DARPA BAA02-21

Here is the second bullet point from the Broad Agency Announcement of DARPA's Cognitive Information Processing Technology Call For Proposals

We didn't get much response from the first bullet posting, and as I said, I may well lose interest if we continue with no response. I will include the preamble for the remainder of these first bullet points.

Just so you know, I rather expected someone to address what constitutes "substantial amounts of appropriately represented knowledge." I am not at all sure that DARPA means high-level ontologies when they state that. THAT's what I thought, but do you all?

I will continue with this exploration for a while, but I would not proceed on this on my own, so nuff said.

Cognitive Information Processing Technology
Proposer Information Pamphlet
Cognitive Information Processing Technology
SOL BAA 02-21 POC: Dr. Ronald J. Brachman, Mr. Zachary J. Lemnios, DARPA/IPTO
E-Mail: baa02-21@darpa.mil
FAX: (703) 741-7804
WEB: http://www.darpa.mil/ipto/Solicitations/index

This is the first set of bullet points I will address wrt to the PIP (Proposer Information Pamphlet) of the BAA. I will post these responses with their own number which relates only to the order in which I post them, and not in any other framework.

This what DARPA is asking for:

"...

The DARPA Information Technology Processing Office (IPTO) is soliciting innovative research proposals in the area of information technology for a new class of cognitive systems that can be characterized simply as follows: a cognitive system is one that, among other things,

* can reason in a variety of ways, using substantial amounts of appropriately represented knowledge;

* can learn from its experiences so that its performance improves as it accumulates knowledge and experience;

* can explain itself and can accept direction;

* can be aware of its own behavior and reflect on its own capabilities; and

* can respond in a robust manner to surprises.

bullet two:

* can learn from its experiences so that its performance improves as it accumulates knowledge and experience;

Well, this one's a pickle and I would like the more expert AI folks among us to clear up the two bottlenecks that will occur in any answer to this stipulation:

1. What is the most widely accepted definition for proof or clear indication of learning?

Is it only that wrong or inaccurate information, e.g. information that is the basis for actions or conclusions which do not yield anticipated outcomes, is corrected and that a series of trials and errors can be conducted by the agent until the most accurate information is adopted as the working hypothesis for a given set of conditions in a given situation, yielding consistent outcomes?

This will result in improved performance statistically, but will it allow the agent to learn and evaluate methods, and arrive at conclusions about how best to conduct the process of learning? That would seem to me to be the more important kind of learning, and I wouldn't mind having some help with that myself.

2. What kind of criteria should an agent have in order to choose the most appropriate experiences to use for models and comparisons when confronted with new situations which call for the use of accumulated knowledge and experience?

Obviously, we need not only to state what criteria we use for 2, but also what definitions, since I suspect there are several, we choose for 1. And, of course, can we know beforehand what definitions DARPA accepts for 1 and what criteria it uses for 2? I suspect we are allowed to ask, but I think such actions should be undertaken by those who will be assembling the actual proposal.

Of course, none of this includes such things as the ability to infer unexpressed information from what is included or left out of information exchanged, or non-verbal behavioral cues, which together comprise one of the more important aspects of cognitive systems.

I'm sure there is much more that needs to be discussed for a satisfactory answer to this bullet point.

I could argue several sides but I get real tired of talking to myself, honest. ;)

Rex Brooks
GeoAddress: 1361-A Addison, Berkeley, CA, 94702 USA, Earth
W3Address: http://www.starbourne.com
Email: rexb@starbourne.com

huml message