Re: [agi] Breaking Solomonoff induction (really)
On 6/21/08, Matt Mahoney [EMAIL PROTECTED] wrote: Eliezer asked a similar question on SL4. If an agent flips a fair quantum coin and is copied 10 times if it comes up heads, what should be the agent's subjective probability that the coin will come up heads? By the anthropic principle, it should be 0.9. That is because if you repeat the experiment many times and you randomly sample one of the resulting agents, it is highly likely that will have seen heads about 90% of the time. That's the wrong answer, though (as I believe I pointed out when the question was asked over on SL4). The copying is just a red herring, it doesn't affect the probability at all. Since this question seems to confuse many people, I wrote a short Python program simulating it: http://www.saunalahti.fi/~tspro1/Random/copies.py Set the number of trials to whatever you like (if it's high, you might want to comment out the A randomly chosen agent has seen... lines to make it run faster) - the ratio will converge to 1:1 on any higher amount of trials. -- http://www.saunalahti.fi/~tspro1/ | http://xuenay.livejournal.com/ Organizations worth your time: http://www.singinst.org/ | http://www.crnano.org/ | http://www.mfoundation.org/ --- 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)
--- On Sun, 6/22/08, Kaj Sotala [EMAIL PROTECTED] wrote: On 6/21/08, Matt Mahoney [EMAIL PROTECTED] wrote: Eliezer asked a similar question on SL4. If an agent flips a fair quantum coin and is copied 10 times if it comes up heads, what should be the agent's subjective probability that the coin will come up heads? By the anthropic principle, it should be 0.9. That is because if you repeat the experiment many times and you randomly sample one of the resulting agents, it is highly likely that will have seen heads about 90% of the time. That's the wrong answer, though (as I believe I pointed out when the question was asked over on SL4). The copying is just a red herring, it doesn't affect the probability at all. Since this question seems to confuse many people, I wrote a short Python program simulating it: http://www.saunalahti.fi/~tspro1/Random/copies.py The question was about subjective anticipation, not the actual outcome. It depends on how the agent is programmed. If you extend your experiment so that agents perform repeated, independent trials and remember the results, you will find that on average agents will remember the coin coming up heads 99% of the time. The agents have to reconcile this evidence with their knowledge that the coin is fair. It is a tricker question without multiple trials. The agent then needs to model its own thought process (which is impossible for any Turing computable agent to do with 100% accuracy). If the agent knows that it is programmed so that if it observes an outcome R times out of N that it would expect the probability to be R/N, then it would conclude I know that I would observe heads 99% of the time and therefore I would expect heads with probability 0.99. But this programming would not make sense in a scenario with conditional copying. Here is an equivalent question. If you flip a fair quantum coin, and you are killed with 99% probability conditional on the coin coming up tails, then, when you look at the coin, what is your subjective anticipation of seeing heads? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
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
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
[agi] Breaking Solomonoff induction (really)
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/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Breaking Solomonoff induction (really)
--- On Fri, 6/20/08, 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. I disagree. AIXI says that the optimal behavior of an agent for maximizing an accumulated reward from a Turing-computable environment while exchanging symbols with it is to guess at each step that the environment is simulated by the shortest program consistent with the interaction so far. AIXI assumes the agent is immortal because it may postpone reward arbitrarily long. The anthropic principle says that events that would have led to the agent's non-existence could not have occurred, and therefore had zero probability. This is inconsistent with Solomonoff induction except in the limit where the agent lives forever. 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. Eliezer asked a similar question on SL4. If an agent flips a fair quantum coin and is copied 10 times if it comes up heads, what should be the agent's subjective probability that the coin will come up heads? By the anthropic principle, it should be 0.9. That is because if you repeat the experiment many times and you randomly sample one of the resulting agents, it is highly likely that will have seen heads about 90% of the time. AIXI is not computable, so humans use the following heuristic approximation: if an experiment is performed N times and a certain outcome occurs R times, and N is large, then the probability of this outcome is estimated to be R/N on the next trial. This is not the right answer in this case. Rather, it is the way we are programmed to think. Remember that probability is just a mathematical approximation of uncertainty. In reality, we cannot assign numerical values to uncertainty. A Solomonoff universal prior is just another model, which depends on a choice of universal Turing machine (and happens to be uncomputable as well). In your example, putting aside an agent is the same as killing it. So the probability of observing 1 correctly converges to 0 for an agent applying the R/N heuristic. AIXI/Solomonoff induction does not apply because this is not a limit case (life expectancy approaching infinity). 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 human upload has more information than the other intelligence because its memories are preserved. Under AIXI it can never guess the simpler model because it would be inconsistent with its past observations. There is no contradiction. 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]