I partially agree and partially disagree with your premise that the structure of concepts must be dynamically created. I believe prototype structures must be predefined, but actual instances can be created on the fly. ~PM +$0.02
> Date: Sat, 4 Apr 2015 10:14:54 -0400 > Subject: Re: [agi] Continuous Reality Checking > From: [email protected] > To: [email protected] > > I think the question of how an effective AGI program can be > constructed is still unanswered. Even supposing (as I do) that you can > start with simple programs (that are not going to be powerful) you > still have to answer the question of how the program can create models > of reality before you can check them. The article, Hybrid Automata for > Formal Modeling and Verification of Cyber-Physical Systems, that you > mentioned looks very interesting and the fact that they are writing > about something that is based on actual experiences is helpful. > However, since their modeling basis does not look entirely relevant, > you have to wonder if they are going to be able to answer the most > important questions that we -should be-asking. > > I believe that most decision models are what I have tried to call > funnel methods. They try to funnel every decision down to an overly > simplistic method. So you can have discrete logic-like methods of > decisions and you can have weighted methods and that is pretty much > it. At this point, when I try to offer my ideas and add a little > competitive ego in order to try to get people to respond, I always end > up with the - well, we already thought of that - kind of response. > That particular response goes along with the sense that our system of > modeling can incorporate that. So in other words, using weighted > reasoning (fuzzy logic, probability and so on) is good enough because > it can incorporate any new ideas about decision making and reasoning > that someone might develop. However, that particular form of hand > waving is where something like your 'reality checking' will inevitably > be reduced to more narrow methods of traditional AI. > > I have tried to point out that the 'structure' of concepts must be > dynamically constructed, and therefore the whole idea that concept > structure is something that can be *entirely* predefined is just not > sound. So, what I am trying to say is, that if you are going to have > your program do some reality checking then it is going to have to be > examining the structural assumptions of the model as well. This leads > to some major questions. > > At any rate, I think it is best to start with a simple semi-agi > program just so I can test an idea like this with simple 'realities'. > > Jim Bromer > > On Sat, Apr 4, 2015 at 7:16 AM, John Rose <[email protected]> wrote: > > > > Where I was going with this was - AGI dynamically builds an internal > > representation of something ... say an observed electromechanical > > machine... and it needs to interact with it and make decisions based on the > > representation. The "consensus reality" I referred to would be the > > commonality of representational model of the system observed by disparate > > intelligent macro-agents, those include people, and the reality checking is > > to also ensure of the safeness of any actions performed based on the > > decisions gleaned from the model. Disregarding however the model exists in > > the AGI mind you still need to interact with the physical world > > periodically to learn and dynamically adjust the model such that prediction > > is maintained and improved. That is even for complex systems > > nondeterministic models like economic models. I’m not for preprogramming > > models but preprogramming the ability to dynamically construct models or > > dynamically emerge models from a more generalized internal systems > > representational capability. > > > > > > > > John > > > > > > > > From: Anastasios Tsiolakidis [mailto:[email protected]] > > Sent: Monday, March 30, 2015 12:00 PM > > To: AGI > > Subject: Re: [agi] Continuous Reality Checking > > > > > > > > > > > > On Sat, Mar 28, 2015 at 11:54 AM, John Rose <[email protected]> wrote: > > > > AGI needs to keep in touch with a consensus reality > > > > > > > > I very much doubt it, I think it is enough for AGI to "prove" it can > > develop increasingly better views of its reality, at its own pace, if an > > AGI passed the Turing test taking an hour for each of its answers I'd be OK > > with that. Also, taking a clue from biology, we do seem to have the fast > > and the slow nervous system as well as engineering solutions like the > > retina and optic nerve being "glued" on the brain rather than the ankles, > > in no way pretending to do reality acquisition, rather defining "real time" > > by their own working tempo, anywhere from 20 to 100 fps. Perhaps the brain > > can do some reading at 10 and 20 fps but none at 100fps, and that's that, > > there seems to be no cache that, after closing one's eyes, allows the > > processing of a backlog. But we do have a backlog mechanism when it comes > > to General Intelligence, we can close our eyes and work through a problem > > domain for seconds, minutes or even longer, with no real-time constraints > > really. > > > > Now, the "social" aspect of intelligence is very important, and it is great > > that we could both in a split second agree about the contents in a room, > > while even 10 years in a room would not be enough to agree on Palestine or > > the Greek debt. For real-world mixed man-machine applications it would be > > important to achieve human-time performance, but not "of the kind" that > > would look like a physics simulation forcing us to tackle some > > analog-to-digital issues in the sense of the article you are quoted. For > > what it's worth I believe that very sparse and cruel representations will > > suffice or even are necessary. I am also convinced that "multiresolution" > > representations will have to be included in any design, analogous to our > > short and long term memory - I am just a bit skeptical of trying to > > "program in" our mind models, for example limiting the artificial short > > term memory to 7 items. > > > > All of the above applies to a kind of "finished product". But during the > > design and evolution of an AGI I have stated elsewhere that indeed one > > could ride the real-time horse, emphasizing the responsiveness of the > > machine. One could, for example, explore the capabilities of one of these > > enormous CUDA cards or FPGAs, acknowledging that you could never respond > > faster than so many nanoseconds (FPGAs being much slower). Then again, the > > optimal "system dimensioning" is the one that includes your sensors and > > actuators, if you can only send commands to your motor 10 times a second, > > then why would you read your body temperature 100 times a second and > > analyze your state 1000 times a second, you are better off running the > > analysis a split second before sending it to the actuator and use 99% of > > your "horsepower" for something else, but what! Biological evolution in > > these cases follows closely the physical constraints of survival and > > reproduction and would not couple a nanosecond brain with a millisecond > > muscle. Our engineering is much more arbitrary and we would find something > > to do with the extra cycles, probably involving longer time scales. However > > that "something" would not have to consider analog domains at all. > > > > > > > > AT > > > > AGI | Archives | Modify Your Subscription > > > > > > > > AGI | Archives | Modify Your Subscription > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/19999924-4a978ccc > Modify Your Subscription: https://www.listbox.com/member/?&; > Powered by Listbox: http://www.listbox.com ------------------------------------------- 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
