Re: [agi] Breaking Solomonoff induction (really)
2008/6/21 Wei Dai [EMAIL PROTECTED]: A different way to break Solomonoff Induction takes advantage of the fact that it restricts Bayesian reasoning to computable models. I wrote about this in is induction unformalizable? [2] on the everything mailing list. Abram Demski also made similar points in recent posts on this mailing list. I think this is a lot stronger objection when you actually implement an implementable variant of Solomonoff Induction (it has started to make me chuckle that a model of induction makes assumptions about the universe that would have to be broken to have it implemented). When you restrict the the memory space of a system a lot more functions become uncomputable with respects to that system. It is not a safe assumption that the world is computable in this restricted notion of computable, i.e. computable with respect to a finite system. Also solomonoff induction ignores any potential physical affects of the computation, as does all probability theory. See section 5 of this attempted paper by me of an formalised example of where things could go wrong. http://codesoup.sourceforge.net/easa.pdf It is not quite an anthropic problem, but it is closely related. I'll tentatively label the observer-world interaction problem. That is the exact nature of the world you see is altered dependent upon the type of system you happen to be. All these are problem with tacit (a la Dennet) representations of beliefs embedded within the Solomonoff induction formalism. Will Pearson --- 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=106510220-47b225 Powered by Listbox: http://www.listbox.com
[agi] Approximations of Knowledge
I just read Abram Demski's comments about Loosemore's, Complex Systems, Artificial Intelligence and Theoretical Psychology, at http://dragonlogic-ai.blogspot.com/2008/03/i-recently-read-article-called-complex.html I thought Abram's comments were interesting. I just wanted to make a few criticisms. One is that a logical or rational approach to AI does not necessarily mean that it would be a fully constrained logical - mathematical method. My point of view is that if you use a logical or a rational method with an unconstrained inductive system (open and not monotonic) then the logical system will, for any likely use, act like a rational-non-rational system no matter what you do. So when, I for example, start thinking about whether or not I will be able to use my SAT system (logical satisfiability) for an AGI program, I am not thinking of an implementation of a pure Aristotelian-Boolean system of knowledge. The system I am currently considering would use logic to study theories and theory-like relations that refer to concepts about the natural universe and the universe of thought, but without the expectation that those theories could ever constitute a sound strictly logical or rational model of everything. Such ideas are so beyond the pale that I do not even consider the possibility to be worthy of effort. No one in his right mind would seriously think that he could write a computer program that could explain everything perfectly without error. If anyone seriously talked like that I would take it as a indication of some significant psychological problem. I also take it as a given that AI would suffer from the problem of computational irreducibility if it's design goals were to completely comprehend all complexity using only logical methods in the strictest sense. However, many complex ideas may be simplified and these simplifications can be used wisely in specific circumstances. My belief is that many interrelated layers of simplification, if they are used insightfully, can effectively represent complex ideas that may not be completely understood, just as we use insightful simplifications while trying to discuss something that is completely understood, like intelligence. My problem with developing an AI program is not that I cannot figure out how to create complex systems of insightful simplifications, but that I do not know how to develop a computer program capable of sufficient complexity to handle the load that the system would produce. So while I agree with Demski's conclusion that, there is a way to salvage Loosemore's position, ...[through] shortcutting an irreducible computation by compromising, allowing the system to produce less-than-perfect results, and, ...as we tackle harder problems, the methods must become increasingly approximate, I do not agree that the contemporary problem is with logic or with the complexity of human knowledge. I feel that the major problem I have is that writing a really really complicated computer program is really really difficult. The problem I have with people who talk about ANNs or probability nets as if their paradigm of choice were the inevitable solution to complexity is that they never discuss how their approach might actually handle complexity. Most advocates of ANNs or probability deal with the problem of complexity as if it were a problem that either does not exist or has already been solved by whatever tired paradigm they are advocating. I don't get that. The major problem I have is that writing a really really complicated computer program is really really difficult. But perhaps Abram's idea could be useful here. As the program has to deal with more complicated collections of simple insights that concern some hard subject matter, it could tend to rely more on approximations to manage those complexes of insight. Jim Bromer --- 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=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Breaking Solomonoff induction (really)
Quick argument for the same point: AIXI is uncomputable, but only considers computable models. The anthropic principle requires a rational entity to include itself in all models that are given nonzero probability. AIXI obviously cannot do so. Such an argument fails for computable approximations of AIXI, however. But they might fail for similar reasons. (Strict AIXI approximations are approximations of an entity that can't reason about itself, therefore any ability to do so is an artifact of the approximation.) On Fri, Jun 20, 2008 at 8:09 PM, Wei Dai [EMAIL PROTECTED] wrote: Eliezer S. Yudkowsky pointed out in a 2003 agi post titled Breaking Solomonoff induction... well, not really [1] that Solomonoff Induction is flawed because it fails to incorporate anthropic reasoning. But apparently he thought this doesn't really matter because in the long run Solomonoff Induction will converge with the correct reasoning. Here I give two counterexamples to show that this convergence does not necessarily occur. The first example is a thought experiment where an induction/prediction machine is first given the following background information: Before predicting each new input symbol, it will be copied 9 times. Each copy will then receive the input 1, while the original will receive 0. The 9 copies that received 1 will be put aside, while the original will be copied 9 more times before predicting the next symbol, and so on. To a human upload, or a machine capable of anthropic reasoning, this problem is simple: no matter how many 0s it sees, it should always predict 1 with probability 0.9, and 0 with probability 0.1. But with Solomonoff Induction, as the number of 0s it receives goes to infinity, the probability it predicts for 1 being the next input must converge to 0. In the second example, an intelligence wakes up with no previous memory and finds itself in an environment that apparently consists of a set of random integers and some of their factorizations. It finds that whenever it outputs a factorization for a previously unfactored number, it is rewarded. To a human upload, or a machine capable of anthropic reasoning, it would be immediately obvious that this cannot be the true environment, since such an environment is incapable of supporting an intelligence such as itself. Instead, a more likely explanation is that it is being used by another intelligence as a codebreaker. But Solomonoff Induction is incapable of reaching such a conclusion no matter how much time we give it, since it takes fewer bits to algorithmically describe just a set of random numbers and their factorizations, than such a set embedded within a universe capable of supporting intelligent life. (Note that I'm assuming that these numbers are truly random, for example generated using quantum coin flips.) A different way to break Solomonoff Induction takes advantage of the fact that it restricts Bayesian reasoning to computable models. I wrote about this in is induction unformalizable? [2] on the everything mailing list. Abram Demski also made similar points in recent posts on this mailing list. [1] http://www.mail-archive.com/agi@v2.listbox.com/msg00864.html [2] http://groups.google.com/group/everything-list/browse_frm/thread/c7442c13ff1396ec/804e134c70d4a203 --- 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/?; Powered by Listbox: http://www.listbox.com --- 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=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Approximations of Knowledge
Jim, On 6/21/08, Jim Bromer [EMAIL PROTECTED] wrote: The major problem I have is that writing a really really complicated computer program is really really difficult. The ONLY rational approach to this (that I know of) is to construct an engine that develops and applies machine knowledge, wisdom, or whatever, and NOT write code yourself that actually deals with articles of knowledge/wisdom. That engine itself will still be a bit complex, so you must write it in Visual Basic or .NET that provides a protected execution environment, and NOT write it in C/C++ that makes it ever so easy to inadvertently hide really nasty bugs. REALLY complex systems may require multi-level interpreters, where a low-level interpreter provides a pseudo-machine on which to program a really smart high-level interpreter, on which you program your AGI. In ~1970 I wrote an ALGOL/FORTRAN/BASIC compiler that ran in just 16K bytes this way. At the bottom was a pseudo-computer whose primitives were fundamental to compiling. That pseudo-machine was then fed a program to read BNF and make compilers, which was then fed a BNF description of my compiler, with the output being my compiler in pseudo-machine code. One feature of this approach is that for anything to work, everything had to work, so once past initial debugging, it worked perfectly! Contrast this with modern methods that consume megabytes and never work quite right. I wrote Dr, Eliza over the course of a year. I developed a daily workflow, that started with answering my email while I woke up. Then came the most creative work - module design. Then came programming, and finally came debugging and testing. Obviously, you need a solid plan to start with to complete such an effort. I spent another year developing my plan, an effort that also involved going to computer conferences and bending the ear of anyone who might have some applicable expertise. On a scale of complexity, Dr. Eliza is MUCH simpler than many of the proposals being made here. However, it does have one salient feature - it actually works in a real-world useful way. The more complex the software, the better the design must be, and the more protected the execution must be. You can NEVER anticipate everything that might go into a program, so they must fail ever so softly. Much of what I have been challenging others on this form for came out of the analysis and design of Dr. Eliza. The real world definitely has some interesting structure, e.g. the figure 6 shape of cause-and-effect chains, and that problems are a phenomenon that exists behind people's eyeballs and NOT otherwise in the real world. Ignoring such things and diving in and hoping that machine intelligence will resolve all (as many/most here seem to believe) IMHO is a rookie error that leads nowhere useful. Steve Richfield --- 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=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Approximations of Knowledge
To be honest, I am not completely satisfied with my conclusion on the post you refer to. I'm not so sure now that the fundamental split between logical/messy methods should occur at the line between perfect approximate methods. This is one type of messiness, but one only. I think you are referring to a related but different messiness: not knowing what kind of environment your AI is dealing with. Since we don't know which kinds of models will fit best with the world, we should (1) trust our intuitions to some extent, and (2) try things and see how well they work. This is as Loosemore suggests. On the other hand, I do not want to agree with Loosemore too strongly. Mathematics and mathematical proof is a very important tool, and I feel like he wants to reject it. His image of an AGI seems to be a system built up out of totally dumb pieces, with intelligence emerging unexpectedly. Mine is a system built out of somewhat smart pieces, cooperating to build somewhat smarter pieces, and so on. Each piece has provable smarts. On Sat, Jun 21, 2008 at 6:54 AM, Jim Bromer [EMAIL PROTECTED] wrote: I just read Abram Demski's comments about Loosemore's, Complex Systems, Artificial Intelligence and Theoretical Psychology, at http://dragonlogic-ai.blogspot.com/2008/03/i-recently-read-article-called-complex.html I thought Abram's comments were interesting. I just wanted to make a few criticisms. One is that a logical or rational approach to AI does not necessarily mean that it would be a fully constrained logical - mathematical method. My point of view is that if you use a logical or a rational method with an unconstrained inductive system (open and not monotonic) then the logical system will, for any likely use, act like a rational-non-rational system no matter what you do. So when, I for example, start thinking about whether or not I will be able to use my SAT system (logical satisfiability) for an AGI program, I am not thinking of an implementation of a pure Aristotelian-Boolean system of knowledge. The system I am currently considering would use logic to study theories and theory-like relations that refer to concepts about the natural universe and the universe of thought, but without the expectation that those theories could ever constitute a sound strictly logical or rational model of everything. Such ideas are so beyond the pale that I do not even consider the possibility to be worthy of effort. No one in his right mind would seriously think that he could write a computer program that could explain everything perfectly without error. If anyone seriously talked like that I would take it as a indication of some significant psychological problem. I also take it as a given that AI would suffer from the problem of computational irreducibility if it's design goals were to completely comprehend all complexity using only logical methods in the strictest sense. However, many complex ideas may be simplified and these simplifications can be used wisely in specific circumstances. My belief is that many interrelated layers of simplification, if they are used insightfully, can effectively represent complex ideas that may not be completely understood, just as we use insightful simplifications while trying to discuss something that is completely understood, like intelligence. My problem with developing an AI program is not that I cannot figure out how to create complex systems of insightful simplifications, but that I do not know how to develop a computer program capable of sufficient complexity to handle the load that the system would produce. So while I agree with Demski's conclusion that, there is a way to salvage Loosemore's position, ...[through] shortcutting an irreducible computation by compromising, allowing the system to produce less-than-perfect results, and, ...as we tackle harder problems, the methods must become increasingly approximate, I do not agree that the contemporary problem is with logic or with the complexity of human knowledge. I feel that the major problem I have is that writing a really really complicated computer program is really really difficult. The problem I have with people who talk about ANNs or probability nets as if their paradigm of choice were the inevitable solution to complexity is that they never discuss how their approach might actually handle complexity. Most advocates of ANNs or probability deal with the problem of complexity as if it were a problem that either does not exist or has already been solved by whatever tired paradigm they are advocating. I don't get that. The major problem I have is that writing a really really complicated computer program is really really difficult. But perhaps Abram's idea could be useful here. As the program has to deal with more complicated collections of simple insights that concern some hard subject matter, it could tend to rely more on approximations to manage those complexes of
Re: [agi] Approximations of Knowledge
Abram, A useful midpoint between views is to decide what knowledge must distill down to, to be able to relate it together and do whatever you want to do. I did this with Dr. Eliza and realized that I had to have a column in my DB that contained what people typically say to indicate the presence of various symptoms (of various cause-and-effect chain links). I now realize that ignorance of the operation of various processes itself is also a condition with its own symptoms, each with their own common expressions of ignorance. OK, so just where was my column going to come from? This information is NOT on the Internet, Wikipedia, etc., yet any expert can rattle this information off in a heartbeat. The only obvious answer was to have experts hand code this information. I am STILL listening to anyone who claims to have another/better way, but I have yet to hear ANY other functional proposal. Of course, this simple realization dooms all of the several efforts now underway to mine the Internet and Wikipedia for knowledge from which to solve problems, yet no one seems to be interested in this simple gotcha, while these doomed efforts continue. I believe that ALL of the ongoing disputes here on this forum are born of a lack of analysis. While the contents of a knowledge base may be very complex and interrelated, the structure of that DB should be relatively simple. This discussion should start with a proposal for structure, and continue as the flaws in that proposal are each identified and addressed. Note in passing that the value of any problem solving system lies in its ability to solve problems with an absolute minimum of information. Hence, systems that require the most information are worth the least, and systems that require all information are completely worthless. Dr. Eliza was designed to operate right at the (currently believed to be) absolute minimum. I completely agree with others here that Dr. Eliza is NOT an AGI as currently envisioned. However, for many of the projected problem-solving functions of a future AGI, it appears to be absolutely unbeatable. People need to either target other functionality for a *useful* future AGI, or else develop designs that won't be predictably inferior to Dr. Eliza. For this, they would do well to fully understand the operation of Dr. Eliza, which should be no problem since it is conceptually pretty simple. Most of the code goes to support speech I/O, the USENET interface, etc., and NOT its core problem solving ability. Steve Richfield === On 6/21/08, Abram Demski [EMAIL PROTECTED] wrote: To be honest, I am not completely satisfied with my conclusion on the post you refer to. I'm not so sure now that the fundamental split between logical/messy methods should occur at the line between perfect approximate methods. This is one type of messiness, but one only. I think you are referring to a related but different messiness: not knowing what kind of environment your AI is dealing with. Since we don't know which kinds of models will fit best with the world, we should (1) trust our intuitions to some extent, and (2) try things and see how well they work. This is as Loosemore suggests. On the other hand, I do not want to agree with Loosemore too strongly. Mathematics and mathematical proof is a very important tool, and I feel like he wants to reject it. His image of an AGI seems to be a system built up out of totally dumb pieces, with intelligence emerging unexpectedly. Mine is a system built out of somewhat smart pieces, cooperating to build somewhat smarter pieces, and so on. Each piece has provable smarts. On Sat, Jun 21, 2008 at 6:54 AM, Jim Bromer [EMAIL PROTECTED] wrote: I just read Abram Demski's comments about Loosemore's, Complex Systems, Artificial Intelligence and Theoretical Psychology, at http://dragonlogic-ai.blogspot.com/2008/03/i-recently-read-article-called-complex.html I thought Abram's comments were interesting. I just wanted to make a few criticisms. One is that a logical or rational approach to AI does not necessarily mean that it would be a fully constrained logical - mathematical method. My point of view is that if you use a logical or a rational method with an unconstrained inductive system (open and not monotonic) then the logical system will, for any likely use, act like a rational-non-rational system no matter what you do. So when, I for example, start thinking about whether or not I will be able to use my SAT system (logical satisfiability) for an AGI program, I am not thinking of an implementation of a pure Aristotelian-Boolean system of knowledge. The system I am currently considering would use logic to study theories and theory-like relations that refer to concepts about the natural universe and the universe of thought, but without the expectation that those theories could ever constitute a sound strictly logical or rational model of everything. Such ideas are so