Re: Questions on Russell's Why Occam paper
I'm not sure I follow your reasoning, but it wouldn't surprise me if the Turing subset of my world has additional constraints - namely the worlds seen by observers whose O(x)'s are prefix machines, not just maps. On Fri, Jun 10, 2005 at 05:56:50PM +0200, Bruno Marchal wrote: Le 10-juin-05, ? 14:59, Patrick Leahy a ?crit : Russell Standish: If the AP applies to the Sims Mark VII, then their reality will be a description containing a body corresponding to their intelligences. They will not be aware of the PC that their description is being generated on. We, who inhabit the world with the PC will not be aware of the countless other PCs, Macs, Xboxes, Eniacs, Turing machines, pebbles in Zen monasteries etc running Sims Mark VII. So the PC itself is actually irrelevant from the internal perspective of the Sims. Well at least we agree on that. No strange loops in this picture, so it is unlike the picture you outline in your paper. Aargh Bad luck! A point where I disagree with both Schmidhuber *and* Standish, at least here apparently. To explain I must assume comp and ... (for one) explicitly the *result* of my thesis. In a nutshell: it is that, if comp is assumed, then the correct law of physics are derivable from comp. (it makes comp testable: derive physics from comp and compare with empirical physics). I will call the physics derived from comp: the comp-physics. Please admit this if only for the sake of the argument. Suppose I build a simulated city with some self-aware entities evolving in that simulated environment. Then Either I simulate the correct comp physics, then apparently the simulated entity cannot know they are simulated by me, but actually this sentence has no meaning, because they are simulated by 2^aleph_0 immaterial stories (constituting arithmetical truth), so it is only in a weak sense that they are failed. (actually it is not even possible to simulate comp physics except in the ridiculous sense of running (really) the universal dovetailer. Or I simulate incorrect comp physics, then the only way we could say the simulated entity are failed is 1) either by killing them (in some absolute way) when they discover discrepancies between the comp-physics they can find by herself and their fake environment. But in that case their story is finite and its measure can be shown equal to 0. Or 2) eithert I keep up correcting the simulation, but then in the limit I don't fail them. Or I limit the cognitive ability of the entities, but then either I will failed to genuinely fail them, or I will make them inconsistent (and here too the measure can be shown equal to zero, and that is related to the non-cul-sac phenomenon). It is an amazing positive consequence of machine's incompleteness that you cannot genuinely failed any (relatively simulated or not) machine having enough introspection power for a very long time. Apparently, In machine's platonia, all lies leads soon or later to a (recognizable) catastroph. Tp prevent falling into an inconsistency, this last conclusion follows from comp, and remember that if comp is correct we cannot know it is correct, and we cannot probably know that all lies leads soon or later to a (recognizable) catastroph. But if you *bet* on comp, you can bet on it! Bruno http://iridia.ulb.ac.be/~marchal/ -- *PS: A number of people ask me about the attachment to my email, which is of type application/pgp-signature. Don't worry, it is not a virus. It is an electronic signature, that may be used to verify this email came from me if you have PGP or GPG installed. Otherwise, you may safely ignore this attachment. A/Prof Russell Standish Phone 8308 3119 (mobile) Mathematics0425 253119 () UNSW SYDNEY 2052 [EMAIL PROTECTED] Australiahttp://parallel.hpc.unsw.edu.au/rks International prefix +612, Interstate prefix 02 pgpREQze8gtH3.pgp Description: PGP signature
Re: Questions on Russell's Why Occam paper
On Fri, Jun 10, 2005 at 01:59:16PM +0100, Patrick Leahy wrote: On Thu, 9 Jun 2005, Russell Standish wrote: Yes, if you think there is a concrete reality in which everything exists (your question of where does the observer live?), then the AP is a tautology. What I meant by where does the observer live, in more formal language, is how do you account for the (apparent) sense data we have?. Perhaps you ask too much. Your question is something along the lines of what explains qualia, or what breathes fire into the equations. I could state baldly that descriptions can conscious. This is no more preposterous than Tegmarks mathematical systems can be conscious. Maybe that satisfies you, maybe not. What can be stated more convincingly is that all experience (including internal conscious experience) is a description, a bunch of data interpreted by an observer. Consequently, the set of all descriptions will contain a description of your current conscious experience. Hence we can make strong claims about appearances - the phenomenon. As to whether there is a Noumenon - I take Laplace's rather agnostic point of view - I have no need of that hypothesis. Perhaps you do. I also have a strong preference for an account where the description of at least our world doesn't privilege one particular observer. In particular, this is hard to square with your insistence that the observer provides the interpretation of each of your bitstring universes. The situation is symmetric with respect to all observers. That is hardly priveleging an observer. As both Hal and I keep trying to emphasise, we are interested in how, or whether, your theory can account for our own existence and the reality (or appearance, if you prefer) that we see around us. So the case of disembodied intelligences is a total a red herring. I don't really care whether these feature in your theory or not, but I do care whether you can account for (apparently) embodied intelligences. Apparently embodied intelligences are part of the space of all descriptions by definition. The anthropic principle accounts for the fact that we observer them, and not (say) flying white rabbits. Instead, one can take the Anthropic Principle as an assertion of the reality we inhabit... Again, you are using a private language... the AP is not regarded as any such assertion by anyone else I've ever heard of. Most people regard their existence as proved by their own subjective experience, not some invented principle. Of course - Decartes and all that. The AP applies to what we see in the world around us, not a proof or otherwise of our own existence. ... and experimentally test it. In all such cases is has been shown to be true, sometimes spectacularly. If we know experimentally the reality we inhabit (?!), which I guess I've just claimed that we do, why do we need a principle to assert it? Likely you mean something completely different, in which case please explain (with examples of said experiments!). The experiments are the usual suspects showing fine tuning of physical parameters in the universe. Tegmark's paper is a good review of the topic. As is Barrow and Tipler's book. snip Quoting me: Then you are implying that the observer can, in a finite time, read and attach meaning to a full (space-time) description of itself, including the act of reading this description and so on recursively. Not at all. Consistency is the only requirement. If the observer goes looking for erself, then e will find erself in the description. It doesn't imply the observer is doing this all the time. I think here we have run into the same inconsistency that you admitted in your discussion with Hal. In your first reply to Hal you assert that the observer O(x) attaches a unique meaning to the description string. Which would imply processing all bits of the string up to the start of the don't care region. A later reply suggest that we should in different contexts assume (a) this and (b) what your paper actually says, i.e. the meanings are updated as further bits are read. Now you have changed this again, and the observer is not (modelled by) a simple mapping but is a free agent who can choose to apply mappings to different regions of the bitstring at will. Yes, as I've further clarified to Hal, O(x) is actually a function of time. One imagines that in general the region of sigificant bits of x expands as a function of (psychological) time. In my defence, the paper was written over a period of 4 years, and the O(x) was a later addition to try to clarify points in section 2. I didn't realise at the time that it introduced some ambiguities into section 3. And even that doesn't actually answer my problem: let's assume the observer *does* go looking for erself. You claim he will find himself, but if the description is *complete* my original problem remains: he will never finish
Re: Questions on Russell's Why Occam paper
Le 10-juin-05, à 14:59, Patrick Leahy a écrit : Russell Standish: If the AP applies to the Sims Mark VII, then their reality will be a description containing a body corresponding to their intelligences. They will not be aware of the PC that their description is being generated on. We, who inhabit the world with the PC will not be aware of the countless other PCs, Macs, Xboxes, Eniacs, Turing machines, pebbles in Zen monasteries etc running Sims Mark VII. So the PC itself is actually irrelevant from the internal perspective of the Sims. Well at least we agree on that. No strange loops in this picture, so it is unlike the picture you outline in your paper. Aargh Bad luck! A point where I disagree with both Schmidhuber *and* Standish, at least here apparently. To explain I must assume comp and ... (for one) explicitly the *result* of my thesis. In a nutshell: it is that, if comp is assumed, then the correct law of physics are derivable from comp. (it makes comp testable: derive physics from comp and compare with empirical physics). I will call the physics derived from comp: the comp-physics. Please admit this if only for the sake of the argument. Suppose I build a simulated city with some self-aware entities evolving in that simulated environment. Then Either I simulate the correct comp physics, then apparently the simulated entity cannot know they are simulated by me, but actually this sentence has no meaning, because they are simulated by 2^aleph_0 immaterial stories (constituting arithmetical truth), so it is only in a weak sense that they are failed. (actually it is not even possible to simulate comp physics except in the ridiculous sense of running (really) the universal dovetailer. Or I simulate incorrect comp physics, then the only way we could say the simulated entity are failed is 1) either by killing them (in some absolute way) when they discover discrepancies between the comp-physics they can find by herself and their fake environment. But in that case their story is finite and its measure can be shown equal to 0. Or 2) eithert I keep up correcting the simulation, but then in the limit I don't fail them. Or I limit the cognitive ability of the entities, but then either I will failed to genuinely fail them, or I will make them inconsistent (and here too the measure can be shown equal to zero, and that is related to the non-cul-sac phenomenon). It is an amazing positive consequence of machine's incompleteness that you cannot genuinely failed any (relatively simulated or not) machine having enough introspection power for a very long time. Apparently, In machine's platonia, all lies leads soon or later to a (recognizable) catastroph. Tp prevent falling into an inconsistency, this last conclusion follows from comp, and remember that if comp is correct we cannot know it is correct, and we cannot probably know that all lies leads soon or later to a (recognizable) catastroph. But if you *bet* on comp, you can bet on it! Bruno http://iridia.ulb.ac.be/~marchal/
Re: Questions on Russell's Why Occam paper
On Thu, Jun 09, 2005 at 01:55:32AM +0100, Patrick Leahy wrote: [Russell Standish wrote]: The AP is a statement that observed reality must be consistent with the observer being part of that reality. Famously, this can be interpreted as either a trivial tautology (Brandon Carter's original intention, I think), or an almost-obviously false principle of necessity (Barrow Tipler's SAP). If you think there's a mystery here it suggests you go for the necessity version, but given your infinite ensemble the tautology would suffice perfectly well. Yes, if you think there is a concrete reality in which everything exists (your question of where does the observer live?), then the AP is a tautology. However, if you are prepared to allow for the possibility that observers exist nowhere, then things are not quite so simple. One can always imagine being the brain-in-the-vat observer a reality which does not contain a body, or a brain, in a vat or anywhere else. Usually in this scenario, the observer will conclude that there must be a body somewhere else, and so concludes that it is inhabiting some kind of virtual reality. However, this implicitly assumes there has to a brain somewhere, and so implies a reality somewhere else for the brain to inhabit. But what if the brain is not required? Obviously, the last conclusion is full blown solipsism, but that is hardly a knock down argument. Instead, one can take the Anthropic Principle as an assertion of the reality we inhabit, and experimentally test it. In all such cases is has been shown to be true, sometimes spectacularly. With the AP, one recovers some of the properties of a concrete reality, without all of it. In particular, Marchal's shared dreaming follows as a consequence, and it contradicts solipsism. You also said: The observer _is_ the interpreter. There may well be more than one observer in the picture, but they'd better agree! Why does this follow? snip It follows from the Anthropic Principle. If O_1 is consistent with its observed reality, and O_2 is consistent with its observed reality, and O_1 observes O_2 in its reality, then O_1 and O_2 must be consistent with each other (at least with respect to their observed realities). Ah. Just to be sure, do you mean that the string the observer attaches meaning to is the one which describes the very same observer? This seems to be implied by your comment above; but you don't say it or clearly imply it in your paper. Then you are implying that the observer can, in a finite time, read and attach meaning to a full (space-time) description of itself, including the act of reading this description and so on recursively. Not at all. Consistency is the only requirement. If the observer goes looking for erself, then e will find erself in the description. It doesn't imply the observer is doing this all the time. Which is impossible, of course. Of course. You also said: I'm not entirely sure I distinguish your difference between external world and internal representation. We're talking about observations here, not models. I'm sure you can distinguish *my* mental representation of the world from your own. Hence if we share a world, and you can't distinguish between that world and your internal representation, then you are not granting equal status to other observers such as me. I'm not sure that is the case. I have a theory of your mind. I get it most economically by observing my own mind, hence I'm self-aware. My theory of the mind says that you are doing the same thing. Isn't this symmetric? You also said (quoting me): My problem is that you are trying to make your observers work at two different levels: as structures within the universes generated (somehow!) by your bitstrings, but also as an interpretive principle for producing meaning by operating *on* the bitstrings. It's a bit like claiming that PCs are built by The Sims. Yes it is a bit like that. Obviously, the Anthropic Principle (or its equivalent) does not work with The Sims. Actually I don't see why not. The existence of The Sims implies a universe compatible with the existence of Sims. But granting this is not so for the sake of the argument, presumably the AP *will* apply to the Sims Mark VII which will be fully self-aware artificial intelligences. If the AP applies to the Sims Mark VII, then their reality will be a description containing a body corresponding to their intelligences. They will not be aware of the PC that their description is being generated on. We, who inhabit the world with the PC will not be aware of the countless other PCs, Macs, Xboxes, Eniacs, Turing machines, pebbles in Zen monasteries etc running Sims Mark VII. So the PC itself is actually irrelevant from the internal perspective of the Sims. But it will still be absurd to claim that the Sims are responsible for construction of PCs (assuming they are not connected to
Re: Questions on Russell's Why Occam paper
Russell Standish writes: On Mon, Jun 06, 2005 at 01:51:36PM -0700, Hal Finney wrote: In particular, if an observer attaches sequences of meanings to sequences of prefixes of one of these strings, then it seems that he must have a domain which does allow some inputs to be prefixes of others. Isn't that what sequences of prefixes would mean? That is, if the infinite string is 01011011100101110111..., then a sequence of prefixes might be 0, 01, 010, 0101, 01011, Does O() apply to this sequence of prefixes? If so then I don't think it is a prefix map. Yes I agree this is vague, and seemingly contradictory. I'm not sure how to make this more precise, but one way to read the paper is to treat observers as prefix maps for section 2 (Occam's razor), and then for section 3 (White Rabbit problem) ignore the prefix property. It could be that the way of making this more precise is to assume observers have some internal state that is constantly updated (a time counter perhaps), so actually going through a sequence of prefix maps in (psychological) time, but at this stage I don't have an answer. Unfortunately I still don't understand this. You agree that it is a seeming contradiction but that doesn't help me to see how to interpret it. Here's an idea. Would it be possible for you to explain how this page is meant to be understood, in an INformal way? Often when people present concepts they do a formal writeup, but if they give a seminar or explanation they will depart from the formalism and explain what is really going on behind the scenes. That's the kind of explanation I think I need. Could you explain how these concepts relate to the actual experiences we have as human observers? What are descriptions and meanings in terms of our sensory and mental experiences? Which descriptions does an observer observe? What are the sequences of prefixes and how do they relate to our day to day lives? What is the point of the equivalence classes and what does that have to do with what we observe? I think an informal explanation of these topics would help me, and perhaps Paddy, to better understand the structure that you formally describe. At this point I am still failing to see how it all relates to my experience of the world as an observer. Thanks - Hal Finney
Re: Questions on Russell's Why Occam paper
[Russell Standish wrote]: The AP is a statement that observed reality must be consistent with the observer being part of that reality. Famously, this can be interpreted as either a trivial tautology (Brandon Carter's original intention, I think), or an almost-obviously false principle of necessity (Barrow Tipler's SAP). If you think there's a mystery here it suggests you go for the necessity version, but given your infinite ensemble the tautology would suffice perfectly well. You also said: The observer _is_ the interpreter. There may well be more than one observer in the picture, but they'd better agree! Why does this follow? snip It follows from the Anthropic Principle. If O_1 is consistent with its observed reality, and O_2 is consistent with its observed reality, and O_1 observes O_2 in its reality, then O_1 and O_2 must be consistent with each other (at least with respect to their observed realities). Ah. Just to be sure, do you mean that the string the observer attaches meaning to is the one which describes the very same observer? This seems to be implied by your comment above; but you don't say it or clearly imply it in your paper. Then you are implying that the observer can, in a finite time, read and attach meaning to a full (space-time) description of itself, including the act of reading this description and so on recursively. Which is impossible, of course. You also said: I'm not entirely sure I distinguish your difference between external world and internal representation. We're talking about observations here, not models. I'm sure you can distinguish *my* mental representation of the world from your own. Hence if we share a world, and you can't distinguish between that world and your internal representation, then you are not granting equal status to other observers such as me. You also said (quoting me): My problem is that you are trying to make your observers work at two different levels: as structures within the universes generated (somehow!) by your bitstrings, but also as an interpretive principle for producing meaning by operating *on* the bitstrings. It's a bit like claiming that PCs are built by The Sims. Yes it is a bit like that. Obviously, the Anthropic Principle (or its equivalent) does not work with The Sims. Actually I don't see why not. The existence of The Sims implies a universe compatible with the existence of Sims. But granting this is not so for the sake of the argument, presumably the AP *will* apply to the Sims Mark VII which will be fully self-aware artificial intelligences. But it will still be absurd to claim that the Sims are responsible for construction of PCs (assuming they are not connected to robot arms etc, for which no analogs exist in your theory). Let alone for them to construct the actual PC on which they are running, as apparently implied by your last message... even robot arms wouldn't help there. Paddy Leahy == Dr J. P. Leahy, University of Manchester, Jodrell Bank Observatory, School of Physics Astronomy, Macclesfield, Cheshire SK11 9DL, UK Tel - +44 1477 572636, Fax - +44 1477 571618
Re: Questions on Russell's Why Occam paper
If we're allowing ourselves a little informality, then I'd appeal to the notion of observer moment. Within any observer moment, a finite number of bits of the bitstrings has been read, and processed by the observer. Since only a finite number of bits have been processed to determine the meaning of reality at that moment, the observer map O(x) is a prefix map. Hence at any point in time the arguments in section 2 of the paper hold. The meaning O(x) could also be called the observer moment. If observer moments are enumerable, one can inject OMs into the set of natural numbers. Observers find themselves embedded in a psychological time. I have not been explicit about exactly what this time is, however I envisage it to probably be what mathematicians call a time scale, which is a closed subset of the real numbers. Time could be continuous, or it could be discrete (eg the set of natural numbers). It could be something else, eg rational numbers or the Cantor set. All of these are example time scales. The exact nature of time is something to be settle later (if possible), but if you are more comfortable witrh discrete time (as many are on this list), then you are welcome to use integers. How this feeds back to our original observer map is that we'd expect the map O(x) to be dependent on time, ie O(t,x). This is consistent with time being psychological. The description or universe x is independent of time. It would correspond to what David Deutsch calls a block universe. Now perhaps section 3 makes some sense. What I call robustness of the observer, ie that observers will not be fooled by a little noise on the line - lions in camouflage are still observed to be lions for instance constrains the form of time evolution of O(t,x). I haven't formalised exactly what this constraint is, but it is something along the lines of continuity of |O^{-1}(t,O(t,x))|, or continuity of the observed complexity of the world. On Wed, Jun 08, 2005 at 09:09:04AM -0700, Hal Finney wrote: Russell Standish writes: On Mon, Jun 06, 2005 at 01:51:36PM -0700, Hal Finney wrote: In particular, if an observer attaches sequences of meanings to sequences of prefixes of one of these strings, then it seems that he must have a domain which does allow some inputs to be prefixes of others. Isn't that what sequences of prefixes would mean? That is, if the infinite string is 01011011100101110111..., then a sequence of prefixes might be 0, 01, 010, 0101, 01011, Does O() apply to this sequence of prefixes? If so then I don't think it is a prefix map. Yes I agree this is vague, and seemingly contradictory. I'm not sure how to make this more precise, but one way to read the paper is to treat observers as prefix maps for section 2 (Occam's razor), and then for section 3 (White Rabbit problem) ignore the prefix property. It could be that the way of making this more precise is to assume observers have some internal state that is constantly updated (a time counter perhaps), so actually going through a sequence of prefix maps in (psychological) time, but at this stage I don't have an answer. Unfortunately I still don't understand this. You agree that it is a seeming contradiction but that doesn't help me to see how to interpret it. Here's an idea. Would it be possible for you to explain how this page is meant to be understood, in an INformal way? Often when people present concepts they do a formal writeup, but if they give a seminar or explanation they will depart from the formalism and explain what is really going on behind the scenes. That's the kind of explanation I think I need. Could you explain how these concepts relate to the actual experiences we have as human observers? What are descriptions and meanings in terms of our sensory and mental experiences? Which descriptions does an observer observe? What are the sequences of prefixes and how do they relate to our day to day lives? What is the point of the equivalence classes and what does that have to do with what we observe? I think an informal explanation of these topics would help me, and perhaps Paddy, to better understand the structure that you formally describe. At this point I am still failing to see how it all relates to my experience of the world as an observer. Thanks - Hal Finney -- *PS: A number of people ask me about the attachment to my email, which is of type application/pgp-signature. Don't worry, it is not a virus. It is an electronic signature, that may be used to verify this email came from me if you have PGP or GPG installed. Otherwise, you may safely ignore this attachment. A/Prof Russell Standish Phone 8308 3119 (mobile) Mathematics0425 253119 () UNSW SYDNEY 2052 [EMAIL PROTECTED] Australia
Re: Questions on Russell's Why Occam paper
Le 06-juin-05, à 22:51, Hal Finney a écrit : I share most of Paddy Leahy's concerns and areas of confusion with regard to the Why Occam discussion so far. I really don't understand what it means to explain appearances rather than reality. Well this I understand. I would even argue that Everett gives an example by providing an explanation of the appearance of a wave collapse from the SWE (Schroedinger Wave equation) and this without any *real*collapse. And I pretend at least that if comp is correct, then the SWE as an *appearance* emerges statistically from the interference of all computations as seen from some inner point of view of the mean universal machine. But, as I pointed a long time ago Russell is hiding (de facto, not intentionally I guess :) many assumptions. There are a lot of derivation of the SWE in the literature, it would be interesting that Russell compares them with its own. My favorite one is the one by Henry and another one by Hardy. Note the incredible derivation of QM from just 5 experiments + a natural principle of simplicity by Julian Swinger in his QM course (taken again by Towsend in its QM textbook). I will give reference once less busy. I agree with Hal and Paddy about the lack of clarity in many passages. Note that my result is infinitely more modest (despite the appearance!). I just prove that if comp is assumed to be correct then a derivation of the SWE *must* exist, without providing it. Well, in the interview of the Lobian machine I do extract some 'quantum logic' from comp, but it is too early to judge if the SWE can be extracted from it. But it should be, in principle, if comp is true. Advantage: I just assume natural numbers and classical logic, I don't assume any geometry or temporality, which for me are really the miraculous things in need to be explained. Bruno It's hard to get my mind around this kind of explanation and what to expect from it. Also the way the Anthropic Principle applies to infinite strings seems extremely vague until we have a clearer picture of how those strings relate to reality. One area I differ: Paddy Leahy writes, quoting Russell: However, as the cardinality of my ensemble is actually c (cardinality of the real numbers), it is quite probably a completely different beast. There you go again with your radical compression. Without the reading I've been doing in the last two weeks, I wouldn't have been able to decode this statement as meaning: 2^\aleph_0 = \aleph_1 (by definition) To assume c = \aleph_1 is the Continuum Hypothesis, which is unprovable (within standard arithmetic). Actually Russell did not bring aleph_1 into the picture at all. All that he referred to was aleph_0 and c which by definition is 2^aleph_0. c is the cardinality of the reals and of infinite bit strings. This is all just definitional. Whether c is the next infinite cardinal after aleph_0 is the Continuum Hypothesis, but that is not relevant here. Another area I had trouble with in Russell's answer was the concept of a prefix map. I understand that a prefix map is defined as a mapping whose domain is finite bit strings such that none of them are a prefix of any other. But I'm not sure how to relate this to the infinite bit strings that are descriptions. In particular, if an observer attaches sequences of meanings to sequences of prefixes of one of these strings, then it seems that he must have a domain which does allow some inputs to be prefixes of others. Isn't that what sequences of prefixes would mean? That is, if the infinite string is 01011011100101110111..., then a sequence of prefixes might be 0, 01, 010, 0101, 01011, Does O() apply to this sequence of prefixes? If so then I don't think it is a prefix map. I want to make it clear by the way that my somewhat pedantic and labored examination of this page is not an attempt to be difficult or stubborn. Rather, I find that by the third page, I don't understand what is going on at all! Even the very first sentence, In the previous sections, I demonstrate that formal mathematical systems are the most compressible, and have highest measure amongst all members of the Schmidhuber ensemble, has me looking to see if I skipped a page! I don't see where this is discussed in any way. So I hope that by pinning down and crystalizing exactly what the first page is claiming, it will help me to see what the more interesting third page is actually able to establish. I think Paddy is in much the same situation. Hal Finney http://iridia.ulb.ac.be/~marchal/
Re: Questions on Russell's Why Occam paper
On Tue, Jun 07, 2005 at 08:29:57AM +0200, Bruno Marchal wrote: Le 06-juin-05, ? 22:51, Hal Finney a ?crit : I share most of Paddy Leahy's concerns and areas of confusion with regard to the Why Occam discussion so far. I really don't understand what it means to explain appearances rather than reality. Well this I understand. I would even argue that Everett gives an example by providing an explanation of the appearance of a wave collapse from the SWE (Schroedinger Wave equation) and this without any *real*collapse. And I pretend at least that if comp is correct, then the SWE as an *appearance* emerges statistically from the interference of all computations as seen from some inner point of view of the mean universal machine. But, as I pointed a long time ago Russell is hiding (de facto, not intentionally I guess :) many assumptions. It would be nice to expose these hidden assumptions. As far as I'm aware, all my assumptions are exposed and upfront, where at least you as a reader can decide if you agree, but there is always the possibility of some that I've missed. There are a lot of derivation of the SWE in the literature, it would be interesting that Russell compares them with its own. My favorite one is the one by Henry and another one by Hardy. The only thing I was aware of by Henry was a derivation of the correspondence principle from gauge invariance in a paper you sent me, something I think that Stenger does better in his book (which is almost published now!). And as for Hardy, I never found his axioms terribly reasonable, unfortunately. Note the incredible derivation of QM from just 5 experiments + a natural principle of simplicity by Julian Swinger in his QM course (taken again by Towsend in its QM textbook). I will give reference once less busy. Sure - I'm not aware of that. I agree with Hal and Paddy about the lack of clarity in many passages. Note that my result is infinitely more modest (despite the appearance!). Hardly infinitely more modest. You start from a slightly different basis (COMP thesis vs all descriptions ensemble), derive the existence of what I assume, and end up not quite where I end up. Perhaps if you adopted Kolmogorov probability axioms, you could get the full QM theory to result. The other things I assume tend to be assumed by you also - COMP = TIME, and I think you assume PROJ. Not sure where your work stand with the Anthropic Principle. I just prove that if comp is assumed to be correct then a derivation of the SWE *must* exist, without providing it. Well, in the interview of the Lobian machine I do extract some 'quantum logic' from comp, but it is too early to judge if the SWE can be extracted from it. But it should be, in principle, if comp is true. Advantage: I just assume natural numbers and classical logic, I don't assume any geometry or temporality, which for me are really the miraculous things in need to be explained. Bruno -- *PS: A number of people ask me about the attachment to my email, which is of type application/pgp-signature. Don't worry, it is not a virus. It is an electronic signature, that may be used to verify this email came from me if you have PGP or GPG installed. Otherwise, you may safely ignore this attachment. A/Prof Russell Standish Phone 8308 3119 (mobile) Mathematics0425 253119 () UNSW SYDNEY 2052 [EMAIL PROTECTED] Australiahttp://parallel.hpc.unsw.edu.au/rks International prefix +612, Interstate prefix 02 pgpQV02YpvmwP.pgp Description: PGP signature
Re: Questions on Russell's Why Occam paper
Le 07-juin-05, à 09:20, Russell Standish a écrit : On Tue, Jun 07, 2005 at 08:29:57AM +0200, Bruno Marchal wrote: Le 06-juin-05, ? 22:51, Hal Finney a ?crit : I share most of Paddy Leahy's concerns and areas of confusion with regard to the Why Occam discussion so far. I really don't understand what it means to explain appearances rather than reality. Well this I understand. I would even argue that Everett gives an example by providing an explanation of the appearance of a wave collapse from the SWE (Schroedinger Wave equation) and this without any *real*collapse. And I pretend at least that if comp is correct, then the SWE as an *appearance* emerges statistically from the interference of all computations as seen from some inner point of view of the mean universal machine. But, as I pointed a long time ago Russell is hiding (de facto, not intentionally I guess :) many assumptions. It would be nice to expose these hidden assumptions. As far as I'm aware, all my assumptions are exposed and upfront, where at least you as a reader can decide if you agree, but there is always the possibility of some that I've missed. OK. it seems to me that (equation 14 at http://parallel.hpc.unsw.edu.au/rks/docs/occam/node4.html ) inline: img75.gif is really presupposing a lot. Where does that come from? It presupposes a space/time geometry, continuity, derivability notion for H, topological notion, etc. To begin with. Bruno http://iridia.ulb.ac.be/~marchal/
Re: Questions on Russell's Why Occam paper
On Tue, Jun 07, 2005 at 10:37:10AM +0200, Bruno Marchal wrote: OK. it seems to me that (equation 14 at http://parallel.hpc.unsw.edu.au/rks/docs/occam/node4.html ) ? In LaTeX, this equation is \frac {d\psi}{d t}={\cal H}(\psi) It supposes time, but not space (TIME postulate). Moreover, it supposes continuous time, but I do suggest in the paper how it might be generalised to other possible timescales. Perhaps it also supposes continuity in time for \psi, although this probably flows from assuming continuity of time. I do not think time is necessarily continuous - I think it is interesting to explore alternative QMs without this assumption. The question is whether this is the most general evolution equation for continuous time, or whether there is some more general equation. Remember, we do have already that \psi is a member of a Hilbert space, so we can write things like: \psi(t')-\psi(t) = ... What do you mean by derivability notion for H, and topological notion? is really presupposing a lot. Where does that come from? It presupposes a space/time geometry, continuity, derivability notion for H, topological notion, etc. To begin with. Bruno http://iridia.ulb.ac.be/~marchal/ -- *PS: A number of people ask me about the attachment to my email, which is of type application/pgp-signature. Don't worry, it is not a virus. It is an electronic signature, that may be used to verify this email came from me if you have PGP or GPG installed. Otherwise, you may safely ignore this attachment. A/Prof Russell Standish Phone 8308 3119 (mobile) Mathematics0425 253119 () UNSW SYDNEY 2052 [EMAIL PROTECTED] Australiahttp://parallel.hpc.unsw.edu.au/rks International prefix +612, Interstate prefix 02 pgpzEx7MGUiyd.pgp Description: PGP signature
Re: Questions on Russell's Why Occam paper
Le 07-juin-05, à 12:28, Russell Standish a écrit : On Tue, Jun 07, 2005 at 10:37:10AM +0200, Bruno Marchal wrote: OK. it seems to me that (equation 14 at http://parallel.hpc.unsw.edu.au/rks/docs/occam/node4.html ) ? In LaTeX, this equation is \frac {d\psi}{d t}={\cal H}(\psi) It supposes time, but not space (TIME postulate). Moreover, it supposes continuous time, Yes but that is a lot of assumptions. Why a linear time capable of being represented by the very special line with the usual topology of the reals? I can imagine many topology on the reals. but I do suggest in the paper how it might be generalised to other possible timescales. yes but if you pretend to derive your equation, I don't understand what you mean by generalizing your conclusion (if only by: I have not derive it and it remains some work to do). Perhaps it also supposes continuity in time for \psi, although this probably flows from assuming continuity of time. Why should a function be continuous just because it is defined on a topological space (which is what I assume you are saying when you say continuity of time). I do not think time is necessarily continuous - I think it is interesting to explore alternative QMs without this assumption. Sure. But again how to talk on derivation then. I mean if someone pretend to derive B from A, then if someone else derive something mùore general than B from A, it is a critic of the assertion that B has been derived from A. If from facts I can derive the murderer is among John and Charles, I am not so interested in knowing the derivation can be generalized into leading that the murderer is among John, Charles, Lee, Bruno and Nicole! The question is whether this is the most general evolution equation for continuous time, or whether there is some more general equation. Absolutely. Remember, we do have already that \psi is a member of a Hilbert space, so we can write things like: OK, but you assume Set theory (that by itself is huge in our context). I show only that you have a preHilbetian space (why should cauchy sequence of vectors converges). \psi(t')-\psi(t) = ... What do you mean by derivability notion for H, and topological notion? Topological notion are needed for talking of continuity (a continuous function is just a function from topological space into a topological space such that the inverse image of open set is an open set).. You assume the familiar topology of reals, complex number, etc. Derivability is a stronger requirement (although some algebraist would introduce many nuances). Someday I will show you make also assumption on consciousness, but that is more subtle, and then all physicist if not almost scientist are doing them when they pretend to solve the consciousness problem like Dennett, or when they put it under the rug (a little bit like Lee in his last posts, I would say). Look Russell, as I said I appreciate your attempt, it is just that, as Hal and Paddy mentionned, there remains quite a lot of work to make it thoroughly communicable. You should really put more clearly your assumptions. You assume a vast part of mathematics, and I would say of physics, mainly with your time postulate and your equation. Compare your work with those I have mentionned (I will give the reference for those you don't have yet). Don't compare it to quickly to mine where the assumptions are made still at a much more basic (logical and arithmetical) level. I assume less than Peano arithmetic. I know I could seem a little bit presomptuous, but nuance would make the post more long and more boring. Hope you don't mind, (actually I would be glad someone criticize the most severely possible my work), Bruno http://iridia.ulb.ac.be/~marchal/
Re: Questions on Russell's Why Occam paper
On Tue, 7 Jun 2005, Russell Standish wrote: Hal dealt with this one already, I notice. 2^\aleph_0 = c. \aleph_1 is something else entirely. d'oh! snip Now an observer will expect to find a SAS in one of the descriptions as a corrolory of the anthropic principle, which is explicitly stated as one of the assumptions in this work. I make no bones about this - I consider the anthropic principle a mystery, not self-evident like many people. Very few supporters of the AP would expect to find a SAS in a bitstring. Until you *specify* a way of interpreting the string, it contains nothing but bits. The observer specifies the interpretation. But the observer is *generated* by the interpretation! Until you have an interpretation, you have no observers. And until you have an observer, you have no interpretation (at least that's how I read the sentence quoted above). How can structures which exist as some sort of pattern inside a bit string (or am I supposed to say in the meaning integer output by some (other) O(x)?) read a separate bitstring which exists as a parallel universe in Platonia? Why should an observer expect to see a token of erself embedded in reality? That is the mystery of the AP. What ARE you talking about? Observer's don't see tokens of themselves... I can see that I have a body - if I look in the mirror I can see a face, eye etc, all of which appear to be under my control. This is a token embedded in my reality that represents me. So you find it a mystery that you have a face, eye etc?? If so, what does the AP have to do with this mystery? Actually, maybe it would clarify things if you said what you mean by the AP; it certainly doesn't seem to be very like the AP that I know about. I'm not sure whether your my reality refers to the external world or your internal representation of it. I guess the latter, otherwise your body would be you, not a token representing you. In particular, any bitstring can be interpreted as any other bitstring by an appropriate map. Hence until you specify an interpreter you are simply not proposing a theory at all. The observer _is_ the interpreter. There may well be more than one observer in the picture, but they'd better agree! Why does this follow? Your observers are maps O(x) from prefix strings to the integers. Why can't you have two inconsistent maps... or rather, how can you possibly avoid such? And since two different maps don't interact at all (how can a mapping interact with another mapping?) each of your observers seems to be sealed in his own little universe. In which case having 1 observer appears to be an unverifiable speculation, which is why I say it seems like solipsism. snip All that is discussed in this paper is appearances - we only try to explain the phenomenon (things as they appear). No attempt is made to explain the noumenon (things as they are), nor do we need to assume that there is a noumenon. Most readers of your paper would take it that you are making a strong ontological proposition, i.e. that the basis of reality is your set of bitstrings. This is the case. Well, if you are making an ontological proposition, you are ipso facto not just explaining appearances. In your model your bitstrings *are* the noumenon (in Kant's terminology). Kant's point was that you can't infer the nature of things in themselves from observation. He didn't say that you can't speculate about their nature, and even guess right, by chance. In effect, this mailing list discusses nothing but the nature of the noumenon. (Kant would probably say this is a waste of time, of course). I think either your terminology or you model has now got very confused. Are your observer TMs the observers (SAS) whose experiences your theory is trying to explain? Yes. In this case where they live is crucial because it defines the environment the SAS find themselves in. Why? An intelligent system is intelligent by virtue of the way it interacts with its environment. Think about the Turing test again: we conclude that the computer is (not) intelligent because of the way it interacts with us. To put it another way, you define these things as observers. This implies something observed. Obviously your model had better account for people (observers) like us observing something like the world we see (where we live), and preferably interacting with other people who are granted equal status your ontology. It is not solipsism, if only for the reason that multiple observers exist in our observed reality. They are all as real as our own consciousness. Bruno Marchal calls this shared dreaming. It seems apt. If that's the *only* reason it's not solipsism, then I would say you just don't have the courage of your convictions. Bruno's shared dreaming sounds very like Leibniz's pre-established harmony, but that only works if you believe in a provident deity (if it ever worked for anyone but Leibniz!).
Re: Questions on Russell's Why Occam paper
On Tue, Jun 07, 2005 at 05:57:17PM +0200, Bruno Marchal wrote: Le 07-juin-05, ? 12:28, Russell Standish a ?crit : On Tue, Jun 07, 2005 at 10:37:10AM +0200, Bruno Marchal wrote: OK. it seems to me that (equation 14 at http://parallel.hpc.unsw.edu.au/rks/docs/occam/node4.html ) ? In LaTeX, this equation is \frac {d\psi}{d t}={\cal H}(\psi) It supposes time, but not space (TIME postulate). Moreover, it supposes continuous time, Yes but that is a lot of assumptions. Why a linear time capable of being represented by the very special line with the usual topology of the reals? I can imagine many topology on the reals. I thought the definition of the reals defined its topology? Perhaps you're using topology somewhat differently. It is a good question as to why time should have a topological dimension of 1, and I admit to not having a good answer to that. All I can say is that computationalism also introduces a time with topological dimension of 1. Tegmark has some arguments as to why the topological dimension is 1, and not any other number, but these are not overly persuasive (the argument depends on properties of 2nd order PDEs, which already carries a greater amount of baggage) I only assume continuity to make contact with standard QM. It is an arbitrary assumption in my opinion, and should rightly be viewed with suspicion. but I do suggest in the paper how it might be generalised to other possible timescales. yes but if you pretend to derive your equation, I don't understand what you mean by generalizing your conclusion (if only by: I have not derive it and it remains some work to do). Of course there remains some work to be done. Perhaps it also supposes continuity in time for \psi, although this probably flows from assuming continuity of time. Why should a function be continuous just because it is defined on a topological space (which is what I assume you are saying when you say continuity of time). Because then it wouldn't be an evolution. For a state \psi(t') to depend on the state \psi(t), there must be a corresponding limit \psi(t')-\psi(t) as t'-t. Of course if t were not continuous, then this condition is no longer necessary. I do not think time is necessarily continuous - I think it is interesting to explore alternative QMs without this assumption. Sure. But again how to talk on derivation then. I mean if someone pretend to derive B from A, then if someone else derive something m?ore general than B from A, it is a critic of the assertion that B has been derived from A. If from facts I can derive the murderer is among John and Charles, I am not so interested in knowing the derivation can be generalized into leading that the murderer is among John, Charles, Lee, Bruno and Nicole! To use your analogy, the situation is more like: Assuming the murderer is John, Charles, Fred or Diana, I have shown that the murderer must be one of John or Charles. However, if we also consider Lee, Bruno and Nicole ... there is more work to be done. The assumption of J,C,F or D corresponds by analogy to the continuity assumption. L,B and N are not continuous - does that sound right? :) The question is whether this is the most general evolution equation for continuous time, or whether there is some more general equation. Absolutely. I think it is, for the reason above. I willing to stand corrected should that not be the case. Remember, we do have already that \psi is a member of a Hilbert space, so we can write things like: OK, but you assume Set theory (that by itself is huge in our context). I show only that you have a preHilbetian space (why should cauchy sequence of vectors converges). Yes I do assume set theory. That is stated. I prove Cauchy sequences converge in the paper. It requires the use of Kolmogorov axiom no. 6. Hence the space of states is a Hilbert space. \psi(t')-\psi(t) = ... What do you mean by derivability notion for H, and topological notion? Topological notion are needed for talking of continuity (a continuous function is just a function from topological space into a topological space such that the inverse image of open set is an open set).. You assume the familiar topology of reals, complex number, etc. Ah yes, you're talking about topological spaces. I just did a quick refresher course on these using Wikipedia. Indeed, when I make the arbitrary continuity assumption of time, it is an assumption that time is a subset of the reals, which has the usual metric and topology defined. When you say you can imagine many topologies on the reals, what you are really saying is that you can imagine sets isomorphic to the reals (1-1 corrspendence), that have many different topologies, eg the trivial or the discrete topology for instance, or anything defined by an arbitrary metric. For instance the set of descriptions which is isomorphic to the
Re: Questions on Russell's Why Occam paper
On Tue, Jun 07, 2005 at 10:15:03PM +0100, Patrick Leahy wrote: Now an observer will expect to find a SAS in one of the descriptions as a corrolory of the anthropic principle, which is explicitly stated as one of the assumptions in this work. I make no bones about this - I consider the anthropic principle a mystery, not self-evident like many people. Very few supporters of the AP would expect to find a SAS in a bitstring. Until you *specify* a way of interpreting the string, it contains nothing but bits. The observer specifies the interpretation. But the observer is *generated* by the interpretation! Until you have an interpretation, you have no observers. And until you have an observer, you have no interpretation (at least that's how I read the sentence quoted above). No the observer is somehow primary. As are the descriptions. If it weren't for the anthropic principle, there would be no connection between the two, and we'd have a genuine brain-in-the-vat. How can structures which exist as some sort of pattern inside a bit string (or am I supposed to say in the meaning integer output by some (other) O(x)?) read a separate bitstring which exists as a parallel universe in Platonia? That is what they do, by definition. Observers observe. Platonia is a collection of all possible observations, hence what observers observe is in Platonia. So you find it a mystery that you have a face, eye etc?? If so, what does the AP have to do with this mystery? Actually, maybe it would clarify things if you said what you mean by the AP; it certainly doesn't seem to be very like the AP that I know about. The AP is a statement that observed reality must be consistent with the observer being part of that reality. I'm not sure whether your my reality refers to the external world or your internal representation of it. I guess the latter, otherwise your body would be you, not a token representing you. I'm not entirely sure I distinguish your difference between external world and internal representation. We're talking about observations here, not models. In particular, any bitstring can be interpreted as any other bitstring by an appropriate map. Hence until you specify an interpreter you are simply not proposing a theory at all. The observer _is_ the interpreter. There may well be more than one observer in the picture, but they'd better agree! Why does this follow? Your observers are maps O(x) from prefix strings to the integers. Why can't you have two inconsistent maps... or rather, how can you possibly avoid such? And since two different maps don't interact at all (how can a mapping interact with another mapping?) each of your observers seems to be sealed in his own little universe. In which case having 1 observer appears to be an unverifiable speculation, which is why I say it seems like solipsism. It follows from the Anthropic Principle. If O_1 is consistent with its observed reality, and O_2 is consistent with its observed reality, and O_1 observes O_2 in its reality, then O_1 and O_2 must be consistent with each other (at least with respect to their observed realities). Most readers of your paper would take it that you are making a strong ontological proposition, i.e. that the basis of reality is your set of bitstrings. This is the case. Well, if you are making an ontological proposition, you are ipso facto not just explaining appearances. In your model your bitstrings *are* the noumenon (in Kant's terminology). Sorry, I don't see this. The bitstrings are phenomena. No noumenon appears. It is possible we are arguing semantic differences only though. In this case where they live is crucial because it defines the environment the SAS find themselves in. Why? An intelligent system is intelligent by virtue of the way it interacts with its environment. Think about the Turing test again: we conclude that the computer is (not) intelligent because of the way it interacts with us. To put it another way, you define these things as observers. This implies something observed. Obviously your model had better account for people (observers) like us observing something like the world we see (where we live), and preferably interacting with other people who are granted equal status your ontology. Does it not? OK, it is not exactly explicit about it, but this situation should appear somewhere in Platonia. It is not solipsism, if only for the reason that multiple observers exist in our observed reality. They are all as real as our own consciousness. Bruno Marchal calls this shared dreaming. It seems apt. If that's the *only* reason it's not solipsism, then I would say you just don't have the courage of your convictions. Bruno's shared dreaming sounds very like Leibniz's pre-established harmony, but that only works if you believe in a provident deity (if it ever worked for anyone but Leibniz!).
Re: Questions on Russell's Why Occam paper
On Mon, 6 Jun 2005, Russell Standish wrote: I am beginning to regret calling the all descriptions ensemble with uniform measure a Schmidhuber ensemble. I think what I meant was that it could be generated by a standard dovetailer algorithm, running for 2^\aleph_0 timesteps. It can't! Timesteps are denumerable, hence this statement is just a contradiction in terms. You better postulate your ensemble without reference to any algorithm to generate it. However, as the cardinality of my ensemble is actually c (cardinality of the real numbers), it is quite probably a completely different beast. There you go again with your radical compression. Without the reading I've been doing in the last two weeks, I wouldn't have been able to decode this statement as meaning: 2^\aleph_0 = \aleph_1 (by definition) To assume c = \aleph_1 is the Continuum Hypothesis, which is unprovable (within standard arithmetic). snip Now an observer will expect to find a SAS in one of the descriptions as a corrolory of the anthropic principle, which is explicitly stated as one of the assumptions in this work. I make no bones about this - I consider the anthropic principle a mystery, not self-evident like many people. Very few supporters of the AP would expect to find a SAS in a bitstring. Until you *specify* a way of interpreting the string, it contains nothing but bits. Why should an observer expect to see a token of erself embedded in reality? That is the mystery of the AP. What ARE you talking about? Observer's don't see tokens of themselves... if anyone (God?) has a 3rd-person/bird's eye view, it is certainly not someone who is included in any particular reality. No way is anything like this implied by the AP. All the AP requires is that there *be* observers/SAS in (real) universes, which is true in our case at least. And now we find not only that the bit string is a description, but it is a complex enough description to describe SAS's? How does that work? The bitstrings are infinite in length. By reading enough bits, they can have arbitrarily complex meanings attached to them. In particular, any bitstring can be interpreted as any other bitstring by an appropriate map. Hence until you specify an interpreter you are simply not proposing a theory at all. snip All that is discussed in this paper is appearances - we only try to explain the phenomenon (things as they appear). No attempt is made to explain the noumenon (things as they are), nor do we need to assume that there is a noumenon. Most readers of your paper would take it that you are making a strong ontological proposition, i.e. that the basis of reality is your set of bitstrings. If this is *not* the case, and you think the bitstrings may be represented in some deeper reality (or maybe are just metaphors), then what is the motivation for your proposal? Why do we need to think about this intermediate layer of bitstrings? The original simplicity goes out the window. BTW I'm with Kant: you can't have an appearance without an underlying reality, even if that is unknowable. Bruno Marchal has a detailed discussion on this in his thesis, and concludes that he has no need for this hypothesis (what he calls the extravagant hypothesis). So the former statement is true :[the description strings are] things that observer TM's observe and map to integers. It is also true that descriptions of self aware observers will appear within the description by the Anthropic Principle. The phenomenon of observerhood is included. However where the observers actually live is not a meaningful question in this framework. I think either your terminology or you model has now got very confused. Are your observer TMs the observers (SAS) whose experiences your theory is trying to explain? In this case where they live is crucial because it defines the environment the SAS find themselves in. If you are not careful your theory becomes effectively that we are all brains in bottles or Leibnizian monads, which is solipsism by another name. Or are your observers the missing interpreters in your theory which give it meaning, and allow us to find (in principle) the SAS within the bitstrings that represent actual observers like us? In this case it's unhelpful to call these meta-entities observers; rather, in effect, they constitute the (meta-)laws of physics. Incidentally, a TM by itself can't generate meaning, as it is only a map from integers to integers. You still have to specify externally how to interpret the code as something more than a mere number. (E.g. in the Turing test the output bits have to be processed into English language text). snip The page then goes on to make some comments about measure applied to universes. Here again I am confused about how to relate it to all that has been descibed. What are the analogs of universes, in this model? Is it descriptions, the infinite bit strings? From what has been
Re: Questions on Russell's Why Occam paper
On Mon, Jun 06, 2005 at 12:06:06PM +0100, Patrick Leahy wrote: On Mon, 6 Jun 2005, Russell Standish wrote: I am beginning to regret calling the all descriptions ensemble with uniform measure a Schmidhuber ensemble. I think what I meant was that it could be generated by a standard dovetailer algorithm, running for 2^\aleph_0 timesteps. It can't! Timesteps are denumerable, hence this statement is just a contradiction in terms. You better postulate your ensemble without reference to any algorithm to generate it. Indeed I do. Only Schmidhuber uses the dovetailer. Hence my regret. ... 2^\aleph_0 = \aleph_1 (by definition) Hal dealt with this one already, I notice. 2^\aleph_0 = c. \aleph_1 is something else entirely. snip Now an observer will expect to find a SAS in one of the descriptions as a corrolory of the anthropic principle, which is explicitly stated as one of the assumptions in this work. I make no bones about this - I consider the anthropic principle a mystery, not self-evident like many people. Very few supporters of the AP would expect to find a SAS in a bitstring. Until you *specify* a way of interpreting the string, it contains nothing but bits. The observer specifies the interpretation. Why should an observer expect to see a token of erself embedded in reality? That is the mystery of the AP. What ARE you talking about? Observer's don't see tokens of themselves... I can see that I have a body - if I look in the mirror I can see a face, eye etc, all of which appear to be under my control. This is a token embedded in my reality that represents me. if anyone (God?) has a 3rd-person/bird's eye view, it is certainly not someone who is included in any particular reality. No way is anything like this implied by the AP. All the AP requires is that there *be* observers/SAS in (real) universes, which is true in our case at least. Sorry - you lost me here ... oh well. And now we find not only that the bit string is a description, but it is a complex enough description to describe SAS's? How does that work? The bitstrings are infinite in length. By reading enough bits, they can have arbitrarily complex meanings attached to them. In particular, any bitstring can be interpreted as any other bitstring by an appropriate map. Hence until you specify an interpreter you are simply not proposing a theory at all. The observer _is_ the interpreter. There may well be more than one observer in the picture, but they'd better agree! snip All that is discussed in this paper is appearances - we only try to explain the phenomenon (things as they appear). No attempt is made to explain the noumenon (things as they are), nor do we need to assume that there is a noumenon. Most readers of your paper would take it that you are making a strong ontological proposition, i.e. that the basis of reality is your set of bitstrings. This is the case. If this is *not* the case, and you think the bitstrings may be represented in some deeper reality (or maybe are just metaphors), then what is the motivation for your proposal? Why do we need to think about this intermediate layer of bitstrings? The original simplicity goes out the window. This latter extrapolation is not the case. BTW I'm with Kant: you can't have an appearance without an underlying reality, even if that is unknowable. I'm not sure Kant says this, but in any case that's not important. I'm with Marchal, who says if there is an underlying reality which is not only unknowable, but also unnecessary to explain phenomena, then why assume that particular hypothesis? It makes no sense. Bruno Marchal has a detailed discussion on this in his thesis, and concludes that he has no need for this hypothesis (what he calls the extravagant hypothesis). So the former statement is true :[the description strings are] things that observer TM's observe and map to integers. It is also true that descriptions of self aware observers will appear within the description by the Anthropic Principle. The phenomenon of observerhood is included. However where the observers actually live is not a meaningful question in this framework. I think either your terminology or you model has now got very confused. Are your observer TMs the observers (SAS) whose experiences your theory is trying to explain? Yes. In this case where they live is crucial because it defines the environment the SAS find themselves in. Why? If you are not careful your theory becomes effectively that we are all brains in bottles or Leibnizian monads, which is solipsism by another name. It is not solipsism, if only for the reason that multiple observers exist in our observed reality. They are all as real as our own consciousness. Bruno Marchal calls this shared dreaming. It seems apt. Or are your observers the missing interpreters in your theory which give it meaning, and
Re: Questions on Russell's Why Occam paper
On Mon, Jun 06, 2005 at 01:51:36PM -0700, Hal Finney wrote: Another area I had trouble with in Russell's answer was the concept of a prefix map. I understand that a prefix map is defined as a mapping whose domain is finite bit strings such that none of them are a prefix of any other. But I'm not sure how to relate this to the infinite bit strings that are descriptions. A prefix map attaches the same output to all strings that share a common finite length prefix. In particular, if an observer attaches sequences of meanings to sequences of prefixes of one of these strings, then it seems that he must have a domain which does allow some inputs to be prefixes of others. Isn't that what sequences of prefixes would mean? That is, if the infinite string is 01011011100101110111..., then a sequence of prefixes might be 0, 01, 010, 0101, 01011, Does O() apply to this sequence of prefixes? If so then I don't think it is a prefix map. Yes I agree this is vague, and seemingly contradictory. I'm not sure how to make this more precise, but one way to read the paper is to treat observers as prefix maps for section 2 (Occam's razor), and then for section 3 (White Rabbit problem) ignore the prefix property. It could be that the way of making this more precise is to assume observers have some internal state that is constantly updated (a time counter perhaps), so actually going through a sequence of prefix maps in (psychological) time, but at this stage I don't have an answer. I want to make it clear by the way that my somewhat pedantic and labored examination of this page is not an attempt to be difficult or stubborn. Even being difficult and stubborn has its place (to help winkle out subtle errors of logic eg), so long as you relax enough at other times to obtain understanding. I appreciate the effort in any case. Rather, I find that by the third page, I don't understand what is going on at all! Even the very first sentence, In the previous sections, I demonstrate that formal mathematical systems are the most compressible, and have highest measure amongst all members of the Schmidhuber ensemble, has me looking to see if I skipped a page! I don't see where this is discussed in any way. This is pretty much a tautology. Formal mathematical systems are a means of compressing data in the form of facts about numbers. If one were to include all such possible compression schemes, rather than just the systems studies by mathematicians to date, one would end up with the set of Turing machines, or equivalently of computable functions. The Occam's razor result clearly relates measure to the amount of compressibility in the description. Perhaps such a view of mathematics is strange. Certainly I find it strange when Stephen Wolfram says mathematics is incapable of understanding complex phenomena, and one should cellular automata instead. To me, cellular automata are just another example of a mathematical system. So I hope that by pinning down and crystalizing exactly what the first page is claiming, it will help me to see what the more interesting third page is actually able to establish. I think Paddy is in much the same situation. Hal Finney I hope so too. -- *PS: A number of people ask me about the attachment to my email, which is of type application/pgp-signature. Don't worry, it is not a virus. It is an electronic signature, that may be used to verify this email came from me if you have PGP or GPG installed. Otherwise, you may safely ignore this attachment. A/Prof Russell Standish Phone 8308 3119 (mobile) Mathematics0425 253119 () UNSW SYDNEY 2052 [EMAIL PROTECTED] Australiahttp://parallel.hpc.unsw.edu.au/rks International prefix +612, Interstate prefix 02 pgpaBNDuiwTTe.pgp Description: PGP signature
Re: Questions on Russell's Why Occam paper
On Fri, Jun 03, 2005 at 04:22:07PM -0700, Hal Finney wrote: Russell Standish recently mentioned his paper Why Occam's Razor which can be found at http://parallel.hpc.unsw.edu.au/rks/docs/occam/ . Among other things he aims to derive quantum mechanics from a Schmidhuber type ensemble. I have tried to read this paper but never really understood it. Here I will try to ask some questions, taking it slowly. On this page, http://parallel.hpc.unsw.edu.au/rks/docs/occam/node2.html , things get started. Russell describes a set of infinite bit strings he calls descriptions. He writes: By contrast to Schmidhuber, I assume a uniform measure over these descriptions -- no particular string is more likely than any other. This surprises me. I thought that Schmidhuber assumed a uniform measure over bit strings considered as programs for his universal computer. So what is the contrast to his work? Nowhere in Schmidhuber (1997) does he propose a measure over the input programs. What he does is is justify the appearance of a universal prior in the set of descriptions by passing the raw data through a reference UTM. Presumably the set of descriptions is sampled uniformly by the observer. In Schmidhuber (2000), the set of descriptions is generated by a machine which has resource bounds. This leads to the notion of speed prior which differs from the universal prior in several important respects. I sometimes refer to the two different ensembles as Schmidhuber I and Schmidhuber II. I am beginning to regret calling the all descriptions ensemble with uniform measure a Schmidhuber ensemble. I think what I meant was that it could be generated by a standard dovetailer algorithm, running for 2^\aleph_0 timesteps. However, as the cardinality of my ensemble is actually c (cardinality of the real numbers), it is quite probably a completely different beast. It is also not generated by a program, Schmidhuber style, it simply is (in the sense of being the simplest set - equivalent to nothing). It seems that the greater contrast is that while Schmidhuber assumed that the bit strings would be fed into a computer that would produce outputs, Russell is taking the bit strings directly as raw data. Quite true. But I am confused about their role. Since some of these descriptions describe self aware substructures... Whoa! This is a big leap for me. First, I am not too happy that mere bit strings have been elevated with the title descriptions. A bit string on its own doesn't seem to have the inherent meaning necessary for it to be considered a description. Many apologies for deploying terminology in a different way to you expect. A description (in my terminology) does not necessarily have meaning. It is simply data. This is in accord with how I use the term in casual English usage too - a description is simply a string of letters, and may or may not be meaningful. Meaning is attached by an observer. Now an observer will expect to find a SAS in one of the descriptions as a corrolory of the anthropic principle, which is explicitly stated as one of the assumptions in this work. I make no bones about this - I consider the anthropic principle a mystery, not self-evident like many people. Why should an observer expect to see a token of erself embedded in reality? That is the mystery of the AP. And now we find not only that the bit string is a description, but it is a complex enough description to describe SAS's? How does that work? The bitstrings are infinite in length. By reading enough bits, they can have arbitrarily complex meanings attached to them. It's especially confusing to read the introductory word since as though this is all quite obvious and need not be explained. To me it is very confusing. Sorry for not going slow enough. The habits of concise expression are hard to shake. The page goes on to identify these SAS's as observers. Now they are mappings, or equivalently Turing Machines, which map finite bit strings to integers. These integers are the meanings of the bit strings. Not equivalently. Not all maps can be represented by a Turing machine, only computable ones. I believe the idea here is that the bit strings are taken as prefixes of the description bit strings in the ensemble. It is as though the observers are observing the descriptions a bit at a time, and mapping them to a sequence of integer meanings. Is that correct? Indeed that is one interpretation. The most important point is that the observer map is a prefix map, in the sense of prefix machines of Algorithmic Information Theory. In reading a bit string one bit at a time, once a meaning is attached to the string, that is the meaning for evermore - the observer cannot change er mind after reading a few more bits. Schmidhuber (2000) deals with machines that do change their mind, so perhaps there is some extension possible in this direction. So here is another confusion about the role of the
Questions on Russell's Why Occam paper
Russell Standish recently mentioned his paper Why Occam's Razor which can be found at http://parallel.hpc.unsw.edu.au/rks/docs/occam/ . Among other things he aims to derive quantum mechanics from a Schmidhuber type ensemble. I have tried to read this paper but never really understood it. Here I will try to ask some questions, taking it slowly. On this page, http://parallel.hpc.unsw.edu.au/rks/docs/occam/node2.html , things get started. Russell describes a set of infinite bit strings he calls descriptions. He writes: By contrast to Schmidhuber, I assume a uniform measure over these descriptions -- no particular string is more likely than any other. This surprises me. I thought that Schmidhuber assumed a uniform measure over bit strings considered as programs for his universal computer. So what is the contrast to his work? It seems that the greater contrast is that while Schmidhuber assumed that the bit strings would be fed into a computer that would produce outputs, Russell is taking the bit strings directly as raw data. But I am confused about their role. Since some of these descriptions describe self aware substructures... Whoa! This is a big leap for me. First, I am not too happy that mere bit strings have been elevated with the title descriptions. A bit string on its own doesn't seem to have the inherent meaning necessary for it to be considered a description. And now we find not only that the bit string is a description, but it is a complex enough description to describe SAS's? How does that work? It's especially confusing to read the introductory word since as though this is all quite obvious and need not be explained. To me it is very confusing. The page goes on to identify these SAS's as observers. Now they are mappings, or equivalently Turing Machines, which map finite bit strings to integers. These integers are the meanings of the bit strings. I believe the idea here is that the bit strings are taken as prefixes of the description bit strings in the ensemble. It is as though the observers are observing the descriptions a bit at a time, and mapping them to a sequence of integer meanings. Is that correct? So here is another confusion about the role of the description bit strings in the model. Are they things that observer TM's observe and map to integers? Or are they places where observers live, as suggested by the Since line quoted above? Or both? Now it gets a little more complicated: Under the mapping O(x), some descriptions encode for identical meanings as other descriptions, so one should equivalence class the descriptions. The problem I have is, O takes only finite bit strings. So technically a description, which is an infinite bit string, does not encode a meaning. What I think is meant here, though, is that two descriptions (i.e. infinite bit strings) will be considered equivalent if for every finite prefix of the strings, the O() mapping is the same. So if we think of O as observing the description bit strings one by one, it will go through precisely the same sequence of integer meanings in each case. Is that right? In particular, strings where the bits after some bit number n are ``don't care'' bits, are in fact equivalence classes of all strings that share the first n bits in common. I think what this considers is a special O() and a special string prefix such that if O sees that particular n-bit prefix, all extensions of that prefix get mapped to the same meaning integer. In that case the condition described in my previous paragraph would be met, and all strings with this n-bit prefix would be equivalent. One can see that the size of the equivalence class drops off exponentially with the amount of information encoded by the string. That seems a little questionable because the size of the equivalence class is infinite in all cases. However I think Russell means to use a uniform measure where the collection of all strings with a particular n-bit prefix have a measure of 1/2^n. It's not clear how well this measure really works or whether it applies to all sets of infinite strings. Under O(x), the amount of information is not necessarily equal to the length of the string, as some of the bits may be redundant. Now we have this new concept of the amount of information which has not previously been defined. This sentence is really hard for me. What does it mean for bits to be redundant? We just discussed strings where all those after bit n are don't care, but this sentence seems to be envisioning other kinds of redundancies. The sum P_O(s) = [sum over p such that O(p)=s of] 2^(-|p|) where |p| means the number of bits of p consumed by O in returning s, gives the size of the equivalence class of all descriptions having meaning s. Boy, that's a tough one now. We consider all bit strings p such that O(p) = s. Now, is this supposed to just be those cases described earlier where the bits after |p| are don't care bits? Or is it all strings p such that