Re: [agi] Complexity is in the system, not the rules themselves
Vladimir Nesov wrote: Richard, These last two messages with replies to Mark's questions clarify your position more clearly than much of your prior writing (although I didn't keep track of later discussions too closely). I think it's important to show in the same example all the controversial aspects: relatively simple rules, use cases where an aspect of global behavior can be modeled by a simple theory (two-body problem, F-14, most of the planets in short term, gliders in GoL), and use cases for the same global system where there is no simple model (n-body problem, Pluto, more general initial state in GoL). Yes, I am coming to the view that this stuff needs to be explained with many examples, if the message has any hope of getting across. More generally, I find it incredibly strange that these complex system ideas cause *so* much consternation. Back in the early 90s I read all about the early history of complex systems research, and it was noticeable that these ideas provoked some extremely strong reactions. People didn't just disagree with the ideas, they were besides themselves with fury. (I am not saying that Mark is doing that, btw, I'm talking about the broader reaction). The funny thing is that I thought all of that was over, and that people now understood what the deal was with complex systems, but what I am finding is that I am fighting exactly the same battles as the earlier folks did, back in the 80s and 90s. But all the same, problems that you describe as complex are just numerical calculation problems. In the case of symbol interaction, initial conditions (rules) are unknown and results are discontinuous, which requires much methodical enumeration to find the rules that give required global behavior, no clever tricks work. I am not quite sure what you mean by this, but my general answer is that it really is not a matter of numerical calculation problems. I think the basic idea that nobody gets (because everyone just dances around the issue) is that if you had a God's-eye perspective, you would be able to plot a distribution graph showing the amount of difficulty that humans have in understanding various kinds of systems (natural and artificial). Looking at that distribution, you would see that most of nature's systems just happen to be clustered in a hump quite close to the origin (i.e. they are low-difficulty), whereas most of the artificial systems in the universe are way, way off up at the high ened of the scale, in a second 'hump'. What this means is that there are two qualitatively different types of systems in the universe: low-difficulty ones, and a second group of extremely high-difficulty ones. But the problem is that people assume that this graph does not have two humps, but is in fact continuous, and that as time goes on our ability to understand systems further up the graph becomes greater. According to that idea science is a relentless march into higher and higher regions of this difficulty-space, so if you came back in a hundred years' time you would find that people are routinely deciphering systems that today require superhuman intellect and then in a thousand years time our elementary school kids will learn String Theory (if it survives that long!), and so on. This view of the relentless march of the human intellect is so strong that I think it comes as a shock to people to be told that things might be different, and that it might be trivially easy to create systems of a certain sort which have a difficulty-level that is so far off the scale that we do not know where to start analysing them, and we may *never* know how to analyze them. But this is exactly what the complex systems idea is about. It really is almost trivial to build an artificial system in which the overall, global behavior of the system is interesting and regular and lawful, but where we have no idea how to prove that this behavior should emerge from the local rules. In that context, it would be a complete misunderstanding to say that the problems that you describe as complex are just numerical calculation problems. If all you mean is that we can simulate them if we want to understand them (the way we simulate the weather in order to predict it), then this is true, but in the context of the problem we have - the problem of building intelligent systems - this fact is of no practical use. Richard Loosemore --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] Complexity is in the system, not the rules themselves
- Original Message Vladimir Nesov [EMAIL PROTECTED] said: In the case of symbol interaction, initial conditions (rules) are unknown and results are discontinuous, which requires much methodical enumeration to find the rules that give required global behavior, no clever tricks work. Vladimir Nesov This is interesting although I had to interpret your comments a little bit. Most symbolic interactions are not numerically commensurate or 'miscible' (so to speak) so they can require a great many methodical operations in order to understand their 'behaviors' at a relatively more global behavior. However, I think this is a problem that can be dealt with. For one thing, many problems can be generalized through various associative methods and that is a trick that does work up to a point. I suspect that we will eventually discover more sophisticated ways to mix a variety of methods of generalization so that even when the combination of data references, reasons, correlations and knowledge of other relations does not produce an easily understandable object of reference, simplifications of the object can be formed using approximations such as approximate correlations. But I think your insight that since interactive symbolic references are not necessarily 'continuous' in some way they may require more elaborate methodologies to understand them is important. Jim Bromer Be a better friend, newshound, and know-it-all with Yahoo! Mobile. Try it now. http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] Complexity is in the system, not the rules themselves
On Thu, May 1, 2008 at 12:04 AM, Jim Bromer [EMAIL PROTECTED] wrote: This is interesting although I had to interpret your comments a little bit. Most symbolic interactions are not numerically commensurate or 'miscible' (so to speak) so they can require a great many methodical operations in order to understand their 'behaviors' at a relatively more global behavior. However, I think this is a problem that can be dealt with. For one thing, many problems can be generalized through various associative methods and that is a trick that does work up to a point. I suspect that we will eventually discover more sophisticated ways to mix a variety of methods of generalization so that even when the combination of data references, reasons, correlations and knowledge of other relations does not produce an easily understandable object of reference, simplifications of the object can be formed using approximations such as approximate correlations. But I think your insight that since interactive symbolic references are not necessarily 'continuous' in some way they may require more elaborate methodologies to understand them is important.Jim Bromer I should add a disclaimer that my comment was based on assumptions that I personally don't agree with, but which I see as underlying Richard's position, which he in turn doesn't really concede... -- Vladimir Nesov [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] Complexity is in the system, not the rules themselves
Vladimir Nesov wrote: On Thu, May 1, 2008 at 12:04 AM, Jim Bromer [EMAIL PROTECTED] wrote: This is interesting although I had to interpret your comments a little bit. Most symbolic interactions are not numerically commensurate or 'miscible' (so to speak) so they can require a great many methodical operations in order to understand their 'behaviors' at a relatively more global behavior. However, I think this is a problem that can be dealt with. For one thing, many problems can be generalized through various associative methods and that is a trick that does work up to a point. I suspect that we will eventually discover more sophisticated ways to mix a variety of methods of generalization so that even when the combination of data references, reasons, correlations and knowledge of other relations does not produce an easily understandable object of reference, simplifications of the object can be formed using approximations such as approximate correlations. But I think your insight that since interactive symbolic references are not necessarily 'continuous' in some way they may require more elaborate methodologies to understand them is important.Jim Bromer I should add a disclaimer that my comment was based on assumptions that I personally don't agree with, but which I see as underlying Richard's position, which he in turn doesn't really concede... And, sadly, none of this really addresses anything I said ;-). Oh well. Richard Loosemore --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
[agi] Complexity is in the system, not the rules themselves
Mark Waser wrote: If I understand Richard correctly, he is assuming that it is necessary to make symbols themselves complex and that each symbol needs his four forces of doom: Memory, Development, Identity, and Non-Linearity. I have no problem with the first three but am not so sure that I agree with the non-linearity. Certainly, the interactions between symbols are non-linear but I believe that they are reasonably bounded -- particularly if you use some intelligent design principles (pun intended). For example, nature re-uses virtually everything -- I have to believe that this applies to cognition as well. Similarly, look at software design patterns (as per Gamma, et. al.). I don't believe at all that rules governing the behavior of inter-symbol interactions are necessarily complex. I believe that inter-symbol interaction will eventually be soluble with a reasonable number of rules (and rules generated from those rules). Just like gravity, the behavior generated by the rules WILL be complex but the rules will not. And just like gravity, there will be more than enough regularity that we will be able to predict and control the stability of inter-symbol interaction *as long as* we understand the rules well enough. More than once in your recent posts, you have said one particular thing that does not make any sense to me, so I need to focus on it. What you said in the above case was I don't believe ... that rules governing the behavior of inter-symbol interactions are necessarily complex. The problem with this statement is that strictly speaking one can never say that the RULES governing a system are complex. Now, before you jump on me (because I have probably made the same mistake), I should say that we sometimes talk that way as a kind of shorthand, but right now we must tread very carefully, so I am going to be very precise: The rules that govern a system are just rules - they are not, by themselves, complex. The SYSTEM can be complex (meaning: you cannot understand global behavior from local rules), but the rules themselves are not complex. But then what can you say about the rules? What you can say about them is whether or not they seem likely to generate complexity. Certain kinds of simple, linear, elegant and separable rules tend not to generate complexity, but other kinds of ugly, tangled rules do tend to generate complexity in the system as a whole. What do I mean by ugly, tangled rules? Well, that was the whole point of me listing the so-called four forces of doom. That list of rule characteristics: - Memory - Development - Identity - Nonlinearity ... is just the sort that tends to make the system as a whole complex. These rules are not complex by themselves, it is just that in our empirical studies of large numbers of experimental systems, putting THOSE kinds of rules in tends to make the system as a whole behave in a complex way. Most often it makes the system just random, of course! But if complexity is going to happen, it is usually because the rules have one or more of those features. So, to illustrate why this is a big deal, look at the quote above: you say that I have no problem with the first three but am not so sure that I agree with the non-linearity. Certainly, the interactions between symbols are non-linear but I believe that they are reasonably bounded... This is not something you can defend: if you think that the rules that govern the behavior of symbols do tend to have three of the four characteristics, then you must expect that the system as a whole will be complex, because this is just an empirical fact. In particular, you cannot say ... the interactions between symbols are non-linear but I believe that they are reasonably bounded Reasonably bounded? That does not buy you anything at all: we can put the tiniest amount of nonlinearity into a system and leave out all the others, and the system still can be complex! Now, it is certainly true that we sometimes utter phrases like the rules governing the system are complex, but that is sloppy, because what we mean is that the rules have enough of these characteristics that the system is complex. I sometimes do this myself, even though I shouldn't, but it is generally harmless. So when you say: Just like gravity, the behavior generated by the rules WILL be complex but the rules will not. ... I have to say that this is a meaningless statement on two counts. First of all, if the rules have some of those four complexity-generating characteristics, the system as a whole will almost always be either complex or random-and-boring. We just do not know of any (many?) examples of a system that has those four in the elements, but where the system as a whole is easily predictable or analysable from its element rules! For anyone to say that they believe that intelligent systems will be the exception is to fly in the face of all empirical
Re: [agi] Complexity is in the system, not the rules themselves
I'm afraid that I'm losing track of your major point but . . . . First off, you are violating your own definition of complexity . . . . You said -- A system is deemed complex if the smallest size of a theory that will explain that system is so large that, for today's human minds, the discovery of that theory is simply not practical. Notice that this definition does not imply that there any such systems in the real world, it just says that *if* the theory size were ever to go off the scale *then* the system would (by definition) be complex. By this definition, gravity is not complex. Yet, below you are arguing that it is, at least, a little bit complex (which seems to be getting more and more analogous to a little bit pregnant :-). Second, you keep whip-sawing between dismissing obviously complex systems like the adaptive aerodynamics of an F-14 as not complex (because whatever that complexity was, it was simple and predictable enough that the control software could actually be written and the complexity could be cancelled out.) and then saying that the least little bit of complexity will make an AI virtually impossible to design. You can't have it both ways. WHY is it that engineers can manage the complexity of high-speed adaptive aerodynamics yet you are absolutely positive that they can't do the same thing for intelligence? I think that the shoe is really on the other foot . . . . what problems *haven't* been eventually solved once we learn enough? True -- intelligence is the mother of all problems, but that doesn't mean that it's too difficult to engineer (like virtually anything else that humankind has put its mind to). - Original Message - From: Richard Loosemore [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Tuesday, April 29, 2008 7:52 PM Subject: [agi] Complexity is in the system, not the rules themselves Mark Waser wrote: If I understand Richard correctly, he is assuming that it is necessary to make symbols themselves complex and that each symbol needs his four forces of doom: Memory, Development, Identity, and Non-Linearity. I have no problem with the first three but am not so sure that I agree with the non-linearity. Certainly, the interactions between symbols are non-linear but I believe that they are reasonably bounded -- particularly if you use some intelligent design principles (pun intended). For example, nature re-uses virtually everything -- I have to believe that this applies to cognition as well. Similarly, look at software design patterns (as per Gamma, et. al.). I don't believe at all that rules governing the behavior of inter-symbol interactions are necessarily complex. I believe that inter-symbol interaction will eventually be soluble with a reasonable number of rules (and rules generated from those rules). Just like gravity, the behavior generated by the rules WILL be complex but the rules will not. And just like gravity, there will be more than enough regularity that we will be able to predict and control the stability of inter-symbol interaction *as long as* we understand the rules well enough. More than once in your recent posts, you have said one particular thing that does not make any sense to me, so I need to focus on it. What you said in the above case was I don't believe ... that rules governing the behavior of inter-symbol interactions are necessarily complex. The problem with this statement is that strictly speaking one can never say that the RULES governing a system are complex. Now, before you jump on me (because I have probably made the same mistake), I should say that we sometimes talk that way as a kind of shorthand, but right now we must tread very carefully, so I am going to be very precise: The rules that govern a system are just rules - they are not, by themselves, complex. The SYSTEM can be complex (meaning: you cannot understand global behavior from local rules), but the rules themselves are not complex. But then what can you say about the rules? What you can say about them is whether or not they seem likely to generate complexity. Certain kinds of simple, linear, elegant and separable rules tend not to generate complexity, but other kinds of ugly, tangled rules do tend to generate complexity in the system as a whole. What do I mean by ugly, tangled rules? Well, that was the whole point of me listing the so-called four forces of doom. That list of rule characteristics: - Memory - Development - Identity - Nonlinearity ... is just the sort that tends to make the system as a whole complex. These rules are not complex by themselves, it is just that in our empirical studies of large numbers of experimental systems, putting THOSE kinds of rules in tends to make the system as a whole behave in a complex way. Most often it makes the system just random, of course! But if complexity is going to happen, it is usually because the rules have one or more of those
Re: [agi] Complexity is in the system, not the rules themselves
Mark Waser wrote: I'm afraid that I'm losing track of your major point but . . . . First off, you are violating your own definition of complexity . . . . You said -- A system is deemed complex if the smallest size of a theory that will explain that system is so large that, for today's human minds, the discovery of that theory is simply not practical. Notice that this definition does not imply that there any such systems in the real world, it just says that *if* the theory size were ever to go off the scale *then* the system would (by definition) be complex. By this definition, gravity is not complex. Yet, below you are arguing that it is, at least, a little bit complex (which seems to be getting more and more analogous to a little bit pregnant :-). No, wait, this is not right. When we talk of 'gravity' we often mean solar-system dynamics (remember, the SYSTEM is what we need to talk about, not the RULES gravity itself is just the low level mechanism, i.e. the rules). But solar systems are a special case of a gravitational system in which most of the behavior is analyzable (thanks to Newton). As I said, if you consider the general case of an n-body system then it is fully and completely complex. But when I say that gravity (by which I mean the solar system) is *partially* complex, I mean that when the orbits are as badly behaved as Pluto's is, the system is unstable. In the specific case of our solar system the presence of Pluto means that the dynamics become grossly unpredictable once in a while. That is the 'little bit of complexity'. That idea of partial complexity is not to be sniffed at. This is not like partial pregnancy. It just means that we can explain some fraction of the system's behavior, but not all of it. Or, that we can explain it most of the time completely, but some of the time the explanation breaks down. Hope that clears it up: I think I have stayed consistent with the original definition I laid out. Second, you keep whip-sawing between dismissing obviously complex systems like the adaptive aerodynamics of an F-14 as not complex (because whatever that complexity was, it was simple and predictable enough that the control software could actually be written and the complexity could be cancelled out.) and then saying that the least little bit of complexity will make an AI virtually impossible to design. You can't have it both ways. WHY is it that engineers can manage the complexity of high-speed adaptive aerodynamics yet you are absolutely positive that they can't do the same thing for intelligence? The problem is that you cannot look all systems that are complex as if they are all complex in the same way. Each one must be examined for its own particular characteristics. In the case of the F-14, it is not the case that there are large numbers of elements that each interact with the others in ways that give rise to the worst kinds of complexity. The plane's designers treat the system as having only TWO components: the plane's body and the environment, with the environment having an unpredictable effect on the plane (it is a noisy signal). They simply [sic!] build a reactive system into the plane so that the plane is measuring the behavior of teh environment and cancelling it out all the time. These two system components do not interact with one another in a way that includes any of the elements that give rise to complexity: the plane's control system just does one simple function, and that is to cancel out all fluctuations to make the plane fly straight. In this case there is a clear situation in which the complexity and randomness is ignored ... the *content* of that randomness and complexity is of no significance whatsoever, because the control system is designed to do only one thing, and that is to cancel it out. A the level of the control system, this F-14 is diabolically simple, it is not complex. The engineers do *not* manage the complexity, they ignore it completely, and pretend that it is just a random signal (in fact, it may be just a random signal with no structure, for all I know: I have not investigated in detail). By contrast (and as I did say before), in the case of intelligence the complexity is happening in and among the very things that cannot be cancelled out. It is impossible to build an AGI by putting into one box all of the symbols and symbol-mechanisms that might possibly cause any complexity, then have an outside system treat that entire box as if it were just a noise source! That outside system would just do its best to pretend that all the stuff going on with the symbols was meaningless noise, cancel it out, and delivery a final output from the system that was well, what? Nothing. The AGI would do nothing intelligent. All symbol activity would have to be cancelled by the compensating mechanism. And within those symbols, there would be huge
Re: [agi] Complexity is in the system, not the rules themselves
Richard, These last two messages with replies to Mark's questions clarify your position more clearly than much of your prior writing (although I didn't keep track of later discussions too closely). I think it's important to show in the same example all the controversial aspects: relatively simple rules, use cases where an aspect of global behavior can be modeled by a simple theory (two-body problem, F-14, most of the planets in short term, gliders in GoL), and use cases for the same global system where there is no simple model (n-body problem, Pluto, more general initial state in GoL). But all the same, problems that you describe as complex are just numerical calculation problems. In the case of symbol interaction, initial conditions (rules) are unknown and results are discontinuous, which requires much methodical enumeration to find the rules that give required global behavior, no clever tricks work. -- Vladimir Nesov [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com