Greetings Duncan,
I looked at the document you attached. I appreciate that you at least
made an attempt to come to an "architecture." In the past (several
years) this list has discussed the AGI architecture several times.
Usually the discussion doesn't last long and quickly diverges to more
specific pet projects.
We might talk more about architecture on the list if we simplified the
architecture to something that any of us could understand. Within broad
categories have more specific discussions. If we were talking about the
architecture of a truck, we would easily recognize the many required
components - wheels, frame, engine, steering mechanism, transmission
etc. Why not start the AGI discussion with components that we can
recognize? To often we hear of exotic, difficult to understand, "flux
capacitors."
Here's my example to give an idea of what I mean by simplified architecture:
First, the AGI has to be an "entity,"- something that has boundaries.
And since the main feature of AGI is intelligence, then the boundary
should reflect "choices." The entity, or unit would need to have limits
to give it any ability to make a choice - it exists in some frame or
context which gives it's choices meaning.
Second, the unit will need to have "cognition." Cognition is an
awareness - the eyes that feed the unit a sense of where it is at.
"Where it is at" is much more that physical or visual, it involves
"situational" sense. (If it can't detect changes, it isn't likely to
make good choices for the moment - wouldn't be intelligent)
Third, the unit will need to have a mechanism to compare several
potential choices. This is where parallel processing can be maximized.
Several processes can be fed the current situational data and each
process can determine if the "time" is right to make the effort to
"push" the choice that is associated with that process. For the record,
I call this mechanism the "promoters" which could exist in the thousands.
You may have noticed that each promoter is pretty "smart." That is,
the promoter seems to have the "secret sauce" to know when the
conditions are ripe to do XYZ. In my view, the promoter is like a
watchdog - not all that smart, but each promoter needs to be "prepped"
with good information. This prepping notion leads to the "hard" parts
of the architecture...
Forth, the unit will need to have an abstraction mechanism to do the
hard calculations that lead to the data given to promoters. This
abstraction unit is hard because it deals with the "general" aspect of
the AGI. Imagine for a moment that one had a very capable "unit" in
that knows the recipe to do hundreds of nifty things. Fine, it could do
many things, but which one is the "best" choice?
The issue for the unit is to determine which of the many recipes to
choose. Granted the promoters are tasked with selecting the recipe for
the moment, but the unit has the task of preparing, and creating,
promoters. The abstractor is the part that builds data about what is
feasible, ethical, moral and beneficial. There is much to be said
about this abstractor component, but let me finish the example
architecture...
Fifth, the unit needs to have an arbitrator mechanism that determines
when the "promoted" choice is of high enough "push" level to exceed the
level of the currently chosen activity. (The arbitrator might also
decide that multiple "tasks" can be executed at the same time.) This
aspect of the unit deals with the strategy or mission that is under
"long term" execution.
And, sixth, there needs to be an underlying messaging mechanism - the
glue that ties the pieces together - mostly electronic message.
So, the example AGI has the following components:
1. a defined entity
2. sensors and world knowledge - cognition
3. promoters
4. abstractor
5. arbitrator
6. messaging or activation ability
The "running" of the machine is simply a cycle that gives cognitions a
chance to be registered, promoters a chance to react to cognitions, and
an arbitrator to determine if the current activity merits special cycles
above the promoted choice.
I hope the "architecture" appears in this writeup.
Stan
(I'll be silent for a few days... going out of town)
On 06/08/2018 05:45 AM, Duncan Murray wrote:
The people in this group have a wide variety of skills and I've seen
some interesting, and confusing approaches. I am a programmer with an
interest in Maths and AI but there are many things posted here that I
don't fully understand.
A lot of people are keeping things closed to others, which is fine -
but if anyone is interested in sharing your work so far, this group
would be a great place to get some initial feedback. Sure it may not
all be positive, but overall we and you may learn something new or
take your ideas in a different direction.
I'll start - first, to prove there are no wrong answers I've attached
my first attempt to work out what an intelligent system might look
like. It looks bit naive, but maybe there is something in there might
help someone else?
Since then, I started working on a framework to try and manage AI
processes, also with limited success.
https://github.com/acutesoftware/AIKIF
So, please - send us links to your projects!
Cheers,
Duncan
*Artificial General Intelligence List
<https://agi.topicbox.com/latest>* / AGI / see discussions
<https://agi.topicbox.com/groups/agi> + participants
<https://agi.topicbox.com/groups/agi/members> + delivery options
<https://agi.topicbox.com/groups> Permalink
<https://agi.topicbox.com/groups/agi/T731509cdd81e3f5f-Me2e6d02cd377344d86af3e1f>
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T731509cdd81e3f5f-M6558358bee68dc9a8ffdabd6
Delivery options: https://agi.topicbox.com/groups
