There is no doubt that complexity, in the sense typically used in dynamical-systems-theory, presents a major issue for AGI systems. Any AGI system with real potential is bound to have a lot of parameters with complex interdependencies between them, and tuning these parameters is going to be a major problem. The question is whether one has an adequate theory of one's system to allow one to do this without an intractable amount of trial and error. Loosemore -- if I interpret him correctly -- seems to be suggesting that for powerful AGI systems no such theory can exist, on principle. I doubt very much this is correct.
-- Ben G On Dec 6, 2007 9:40 AM, Ed Porter <[EMAIL PROTECTED]> wrote: > Jean-Paul, > > Although complexity is one of the areas associated with AI where I have less > knowledge than many on the list, I was aware of the general distinction you > are making. > > What I was pointing out in my email to Richard Loosemore what that the > definitions in his paper "Complex Systems, Artificial Intelligence and > Theoretical Psychology," for "irreducible computability" and "global-local > interconnect" themselves are not totally clear about this distinction, and > as a result, when Richard says that those two issues are an unavoidable part > of AGI design that must be much more deeply understood before AGI can > advance, by the more loose definitions which would cover the types of > complexity involved in large matrix calculations and the design of a massive > supercomputer, of course those issues would arise in AGI design, but its no > big deal because we have a long history of dealing with them. > > But in my email to Richard I said I was assuming he was not using this more > loose definitions of these words, because if he were, they would not present > the unexpected difficulties of the type he has been predicting. I said I > though he was dealing with more the potentially unruly type of complexity, I > assume you were talking about. > > I am aware of that type of complexity being a potential problem, but I have > designed my system to hopefully control it. A modern-day well functioning > economy is complex (people at the Santa Fe Institute often cite economies as > examples of complex systems), but it is often amazingly unchaotic > considering how loosely it is organized and how many individual entities it > has in it, and how many transitions it is constantly undergoing. Unsually, > unless something bangs on it hard (such as having the price of a major > commodity all of a sudden triple), it has a fair amount of stability, while > constantly creating new winners and losers (which is a productive form of > mini-chaos). Of course in the absence of regulation it is naturally prone > to boom and bust cycles. > > So the system would need regulation. > > Most of my system operates on a message passing system with little concern > for synchronization, it does not require low latencies, most of its units, > operate under fairly similar code. But hopefully when you get it all > working together it will be fairly dynamic, but that dynamism with be under > multiple controls. > > I think we are going to have to get such systems up and running to find you > just how hard or easy they will be to control, which I acknowledged in my > email to Richard. I think that once we do we will be in a much better > position to think about what is needed to control them. I believe such > control will be one of the major intellectual challenges to getting AGI to > function at a human-level. This issue is not only preventing runaway > conditions, it is optimizing the intelligence of the inferencing, which I > think will be even more import and diffiducle. (There are all sorts of > damping mechanisms and selective biasing mechanism that should be able to > prevent many types of chaotic behaviors.) But I am quite confident with > multiple teams working on it, these control problems could be largely > overcome in several years, with the systems themselves doing most of the > learning. > > Even a little OpenCog AGI on a PC, could be interesting first indication of > the extent to which complexity will present control problems. As I said if > you had 3G of ram for representation, that should allow about 50 million > atoms. Over time you would probably end up with at least hundreds of > thousand of complex patterns, and it would be interesting to see how easy it > would be to properly control them, and get them to work together as a > properly functioning thought economy in what ever small interactive world > they developed their self-organizing pattern base. Of course on such a PC > based system you would only, on average, be able to do about 10million > pattern to pattern activations a second, so you would be talking about a > fairly trivial system, but with say 100K patterns, it would be a good first > indication of how easy or hard agi systems will be to control. > > Ed Porter > > -----Original Message----- > From: Jean-Paul Van Belle [mailto:[EMAIL PROTECTED] > Sent: Thursday, December 06, 2007 1:34 AM > To: agi@v2.listbox.com > > Subject: RE: [agi] None of you seem to be able ... > > Hi Ed > > You seem to have missed what many A(G)I people (Ben, Richard, etc.) mean by > 'complexity' (as opposed to the common usage of complex meaning difficult). > It is not the *number* of calculations or interconnects that gives rise to > complexity or chaos, but their nature. E.g. calculating the eigen-values of > a n=10^10000 matrix is *very* difficult but not complex. So the large matrix > calculations, map-reduces or BleuGene configuration are very simple. A > map-reduce or matrix calculation is typically one line of code (at least in > Python - which is where Google probably gets the idea from :) > > To make them complex, you need to go beyond. > E.g. a 500K-node 3 layer neural network is simplistic (not simple:), > chaining only 10K NNs together (each with 10K input/outputs) in a random > network (with only a few of these NNs serving as input or output modules) > would produce complex behaviour, especially if for each iteration, the input > vector changes dynamically. Note that the latter has FAR FEWER interconnects > i.e. would need much fewer calculations but its behaviour would be > impossible to predict (you can only simulate it) whereas the behaviour of > the 500K is much more easily understood. > BlueGene has a simple architecture, a network of computers who do mainly the > same thing (e.g the GooglePlex) has predictive behaviour, however if each > computer acts/behaves very differently (I guess on the internet we could > classify users into a number of distinct agent-like behaviours), you'll get > complex behaviour. It's the difference in complexity between a 8Gbit RAM > chip and say an old P3 CPU chip. The latter has less than one-hundredth of > the transistors but is far more complex and displays interesting behaviour, > the former doesn't. > > Jean-Paul > >>> On 2007/12/05 at 23:12, in message > <[EMAIL PROTECTED]>, > "Ed Porter" <[EMAIL PROTECTED]> wrote: > > Yes, my vision of a human AGI would be a very complex machine. Yes, > > a lot of its outputs could only be made with human level reasonableness > > after a very large amount of computation. I know of no shortcuts around > the > > need to do such complex computation. So it arguably falls in to what you > > say Wolfram calls "computational irreducibility." > > But the same could be said for any of many types of computations, > > such as large matrix equations or Google's map-reduces, which are > routinely > > performed on supercomputers. > > So if that is how you define irreducibility, its not that big a > > deal. It just means you have to do a lot of computing to get an answer, > > which I have assumed all along for AGI (Remember I am the one pushing for > > breaking the small hardware mindset.) But it doesn't mean we don't know > how > > to do such computing or that we have to do a lot more complexity research, > > of the type suggested in your paper, before we can successfully designing > > AGIs. > [...] > > Although it is easy to design system where the systems behavior > > would be sufficiently chaotic that such design would be impossible, it > seems > > likely that it is also possible to design complex system in which the > > behavior is not so chaotic or unpredictable. Take the internet. > Something > > like 10^8 computers talk to each other, and in general it works as > designed. > > Take IBM's supercomputer BlueGene L, 64K dual core processor computer each > > with at least 256MBytes all capable of receiving and passing messages at > > 4Ghz on each of over 3 dimensions, and capable of performing 100's of > > trillions of FLOP/sec. Such a system probably contains at least 10^14 > > non-linear separately functional elements, and yet it works as designed. > If > > there is a global-local disconnect in the BlueGene L, which there could be > > depending on your definition, it is not a problem for most of the > > computation it does. > > -- > > Research Associate: CITANDA > Post-Graduate Section Head > Department of Information Systems > Phone: (+27)-(0)21-6504256 > Fax: (+27)-(0)21-6502280 > Office: Leslie Commerce 4.21 > > > ----- > This list is sponsored by AGIRI: http://www.agiri.org/email > To unsubscribe or change your options, please go to: > http://v2.listbox.com/member/?&; > > ----- > This list is sponsored by AGIRI: http://www.agiri.org/email > To unsubscribe or change your options, please go to: > http://v2.listbox.com/member/?&; ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=73163220-1ae588
