The first step is carving up the world so that you can talk about it easily.
So we need several categories. In the Premise language, we could specify them
as follows:
let percept :id :nameend
let concept :id :nameend
let action :id :nameend
let solution :id :before { } :steps { } :after { } end
let prediction :id :before { } :after { }end
let episode :id :before { } :during { } :after { }end
In PAM P2 it's actually more complicated than this but the idea is that you are
working at the level of solutions most of the time, rather than at the level of
actions. Actions are primitive. If actions are encountered they are matched to
solutions which contain them, or combined to form new solutions. Similarly for
percepts (external elements) or concepts (internally synthesized elements).
Percepts and concepts are matched to predictions and solutions.
Predictions and solutions are practically the same thing. The antecedent
(:before slot) of a prediction is equivalent to both the antecedent and steps
slots (:before and :steps) of a solution.
Jim Bromer: "So how could an AGI program determine that the action produced a
kind of reaction(that can be used as an intrinsic goal)?"
The answer, is through episode analysis (or episode mining).
So in PAM P2 when an action is taken by the Executor component, all
subsequently activated percepts and concepts form episodes. Episodes are the
basis of solutions and predictions. Episodes can be nested and sequenced. So
you capture episodes at multiple levels and then study them for
derivingpredictions, solutions, etc.
Very simple. This is one approach. Doubtless there are many others.
~PM Date: Sun, 8 Dec 2013 07:46:51 -0500
Subject: Re: [agi] Key Structural Insights Should Provide Some Weight to a Good
Solution
From: [email protected]
To: [email protected]
I was thinking that if it was discovered that a particular action usually
produced a response then that would knowledge would be turned into an implicit
goal (if the expectation usually occurred after the response then the matched
expectation would be reinforcing). This would be the "prediction" of AI lore.
But, the problem is that most actions do not usually produce the exact same
response or reaction that had been seen before. So how could an AGI program
determine that the action produced a kind of reaction (that could be used as an
intrinsic goal). My answer to that question is that structural keys would help
in identifying more complex situations that could be used to determine whether
the action produced a kind of response. For example, your remarks are examples
of a response which is directly related to the subject matter of my original
comment. So now I have a higher expectation that you would be able to better
understand my understanding of awareness which is a kind of meta knowledge
about a subject matter - and about how it can be used in an activation by a
program.
On Sat, Dec 7, 2013 at 4:51 PM, Piaget Modeler <[email protected]>
wrote:
One approach to this conundrum is to assume that "awareness" = activation.
When a concept in the memory is activated, the AGI is aware of it. The
question
then becomes how to activate these concepts. How to bring about the activation.
Then the AGI must set a goal to achieve the desired concept's activation and
formulate or retrieve a solution which achieves the goal.
When actions are taken towards a goal, then reflective processes must evaluate
the success or failure of the attempt. In PAM-P2, attempts have a time frame.
When a particular solution is attempted, if the goal is not achieved within the
time
frame, the solution has failed, otherwise it has succeeded. Regulatory
processes will reinforce and recombine successful solutions, and correct failed
solutions.
Moreover, compensatory processes will act upon the solution to achieve a
desired state.
All this is using intrinsic reinforcement, rather than extrinsic reinforcement.
~PM
-----------
> > On 12/07/2013 03:16 AM, Jim Bromer wrote:
> >>
> >> One of the problems is how do you get an AGI program to be 'aware'
> >> that it has found an appropriate solution to 'understanding' a
> >> situation without some kludgy method of external reinforcement? I
> >> believe that key structural insights may play an important role in
> >> this process. I assume that most learning takes place through an
> >> incremental process of accumulating small pieces of insight or
> >> know-how.
> >>
>
AGI | Archives
| Modify
Your Subscription
--
Jim Bromer
AGI | Archives
| Modify
Your Subscription
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
