Does anyone else agree that, as a piece of expository writing, that document sucks bigtime?
111 pages! I got through about 25 and I was wondering why Eliezer thought I needed to hear about how his four friends had decided when presented with the Newcomb’s soda problem and how some people refer to this problem as Solomon’s problem. So, I decided to skim ahead until he started talking about TDT. So I skimmed and skimmed.
Finally, I got to section 14, entitled “The timeless decision procedure”. “Aha!”, I think. “Finally.” The first paragraph consists of a very long and confusing sentence which at least seems to deal with the timeless decision procedure.
The timeless decision procedure evaluates expected utility conditional upon the output of an abstract decision computation—the very same computation that is currently executing as a timeless decision procedure—and returns that output such that the universe will possess maximum expected utility, conditional upon the abstract computation returning that output.
It might be easier to understand if expressed as an equation or formula containing, you know, variables and things. So I read on, hoping to find something I can sink my teeth into. But then the second paragraph begins:
I delay the formal presentation of a timeless decision algorithm because of some significant extra steps I wish to add …
and closes with
Before adding additional complexities, I wish to justify this critical innovation from first principles.
As far as I can tell, the remainder of this section entitled “The timeless decision procedure” consists of this justification, though not from first principles, but rather using an example. And it doesn’t appear that Eliezer ever gets back to the task of providing a “formal presentation of a timeless decision algorithm”.
So, I skip forward to the end, hoping to read the conclusions. Instead I find:
This manuscript was cut off here, but interested readers are suggested to look at these sources for more discussion:
Followed by a bibliography containing one entry—A chapter from a 1978 collection of articles on applications of decision theory.
″...was cut off here …”? Give me a break!
Let me know when you get it down to a dozen pages or so.
ETA: A cleaned up copy of the paper exists with a more complete bibliography and without the “manuscript was cut off here” closing.
The first paragraph consists of a very long and confusing sentence which at least seems to deal with the timeless decision procedure.
The timeless decision procedure evaluates expected utility conditional upon the output of an abstract decision computation—the very same computation that is currently executing as a timeless decision procedure—and returns that output such that the universe will possess maximum expected utility, conditional upon the abstract computation returning that output.
I think this needs rewriting so it doesn’t sound so circular—and only mentions the word “conditional” once.
It seems to me that we can just say that it maximises utility—while maintaining an awareness that there may be other agents running its decision algorithm out there, in addition to all the other things it knows.
I think the stuff about “conditional upon the abstract computation returning that output” is pretty-much implied by the notion of utility maximisation.
I guess this is a case of “different strokes for different folks”. I will point out that it is fairly traditional for technical communication to contain formulas, equations, and/or pseudo-code. I believe the assumption behind this tradition is that such formal means of presentation are often clearer than expository text.
I will point out that it is fairly traditional for technical communication to contain formulas, equations, and/or pseudo-code.
I am aware of the tradition. Yes, Eliezer’s piece does not include any semblance of technical rigour.
I believe the assumption behind this tradition is that such formal means of presentation are often clearer than expository text.
There is a reason the formal presentations include accompanying explanations. The mathematics for this kind of thing would be nigh incoherent and quite possibly longer than a verbal description. Expository text is utterly critical.
Incidentally, I have almost no doubt that “might be easier to understand” is not your real reason for demanding “you know, variables and things”. Some of your real reasons may actually be better in this instance.
Thanks for the link! I just read it all. The good: it’s very, very smooth reading—I know how well Eliezer can write, and even I was surprised at the quality—and it has some very lucid explanations of tricky matters (like why Pearlean causality is useful). The bad: it’s kinda rambling, contains many standard sci-fi LW arguments that feel out of place in a philosophy paper, and it doesn’t make any formal advances beyond what we already know here (I’d hoped to see at least one). The verdict: definitely read the first half if you’re confused about this whole “decision theory” controversy, it’ll get you unconfused in a pinch. Take the second half with a grain of salt because it’s still very raw (unmixed metaphor award!)
I’m currently reading through the document, and yes, it definitely should. The present format is an unprofessional-looking eyesore, and the references are presented in a weird, clumsy, and inconsistent way. Using Latex/Bibtex would solve both problems easily and effectively.
(Personally, I can’t fathom why anyone capable of grasping the notion of a markup language would ever want to write a document longer than five pages in Word instead of Latex.)
From a list of warning signs of a FAIL in an attempt to solve a famous problem:
The paper doesn’t build on (or in some cases even refer to) any previous work.
The paper wastes lots of space on standard material.
I would disagree that this paper doesn’t build on or take notice of previous work. It takes note of EDT and CDT and quite properly puts the focus on the point of departure of this work—specifically, the handling of contrafactuals. I’m quite happy with that aspect of the paper. My complaint was (8) that it wasted far too much space doing it.
And, perhaps as a result of wasting so much time and space in preparation, it never reached its proper conclusion.
Also, it is not completely clear the Aronson’s list of warning signs really applies here. Eliezer is not solving a famous problem here. Most non-philosophers don’t think that a problem even exists. So, he does have to provide an explanation of why TDT is needed. Just not so much explanation.
Also, it is not completely clear the Aronson’s list of warning signs really applies here.
Nor do I, and I would in any case suggest that some of them are screened off. There’s only so many times you can count ‘non-conventional’ as evidence.
I incidentally found some of the extra explanation handy purely as revision of various topics that it hadn’t particularly occurred to me were relevant.
And, perhaps as a result of wasting so much time and space in preparation, it never reached its proper conclusion.
I do hope someone goes ahead and finishes it. Including things like writing out that bibliography at the end and writing up the maths.
I must say I’m disappointed by the lack of rigor. On the other hand, I’m slightly relieved that he didn’t beat me to any of the stuff in the decision theory document I’m writing myself. So far, I have yet to see any formalization of decision theory that I would consider usable, other my own unfinished one.
I notice there seems to be an issue with the bibliography—there’s only one entry in it, but I’ve found at least one other citation in the text (Judea Pearl’s Causality cited on page 58) that’s not there. Are there any good collections of decision theory paper links out there?
If you have new formal arguments about decision theory, it would be much more useful to me (and others, I think) if you just posted them here in their current state instead. Or emailed them to the interested people.
I’m approaching decision theory from from the perspective compilers approach optimizations: no approach is guaranteed to work always, but each one comes with a list of preconditions that you can check. I’m also summarizing some of the relevant work from compilers: automatic provably correct simplification, translation between forms, and a handy collection of normal forms to translate into.
For CDT, the precondition is a partial ordering over observation sets passed to the strategy such that the world program calls the strategy with observation sets only in increasing order, and there are finitely many possible observation sets. Then you can translate the program into continuation-passing style, and enumerate the possible invocations of the strategy function and their ordering. The last one in the order is guaranteed to have a continuation with no further invocations of the strategy function, which means you can try each possibility, simulate the results, and use that to determine the best answer. Then you can look at the second-to-last invocation, substitute the best answer to the last invocation into the continuation, and repeat; and so on for the set of all invocations to the strategy function. This works because you have a guarantee that when you compute your current position within the world-program and come up with a probability distribution over states to determine where you are, and then look at future continuations, changing result of any invocations of the strategy in those continuations does not affect the probability distribution over states.
I also have an example of a formalized decision-theory problem for which no optimal answer exists: name a number and that number is your utility. A corollary is that no decision theory can always give optimal answers, even given infinite computing power. This can be worked around by applying size bounds in various places.
I’m also drawing distinctions between strategies and decision theories (a strategy is an answer to one problem, a decision theory is an approach to generating strategies from problems); and between preference and utility (a preference is a partial order over outcomes; a utility function is a total order over outcomes where the outcomes are complete probability distributions, plus a linearity requirement).
By that, do you mean that it sounds wrong, or that it sounds confused? If the former, I may need to reconsider; if the latter, I’m unsurprised because it’s much too short and doesn’t include any of the actual formalization. (That was not an excerpt from the draft I’m writing, but an attempt to summarize it briefly. I don’t think I did it justice.)
Ok, in that case I’m inclined to think that impression is just an artifact of how I summarized it, since my summary didn’t address the questions, but the longer paper I’m working on does, albeit only after building up proof and formalization techniques, which are the main focus.
As far as I know, there are no cases where UDT suggests a decision and disagrees with mine. The differences are all in cases where UDT alone can’t be used to reach a decision.
I notice that the ideal causal diagram used in Part 2 (and based on pearls) is isomorphic to an example I use to teach CLIP, once you apply the substitution:
sprinkler on → a paperclip truck has over turned rain → a clippy has haphazardly used up a lot of metal wire sidewalk wet → paperclips are scattered across the ground sidewalk slippery → many paperclips need to be moved to the safe zone
I’m glad to have this to read. I was surprised to find many examples and arguments that EY hadn’t given before (or at least formalized this way). I liked the Newcomb’s soda problem in particular. I had been worried that EY had presented enough of his TDT justification that someone could “scoop” him, but there’s a lot more depth to it. (Anyone up-to-date on the chance that he could get a PhD just for this?)
And I also appreciated that the modified the smoking lesion problem to be one where people aren’t distracted by their existing knowledge of smoking, and that this was the reason for transforming the example.
I read up to ~p. 35, and I think I have a better understanding now of the relevance of time consistency and how it varies across examples.
That said, I agree with the others who say it could use some mroe polish.
Yeah, Newcomb’s Soda and Solomon’s Problem are really interesting! If I faced the problems right now, I would one-box, skip the gum, and eat chocolate ice cream, because those choices put me in the groups that win, so I guess that classifies me as an evidentialist. At the same time, I haven’t reasoned out these conclusions thoroughly—I can’t argue formally against causal reasoning or dominance, or formally for evidentialism.
Looks like I have some more reading to do before I get this resolved.
Edit:
All right, I think I can clarify my positions. I would not say that I am choosing based on evidential reasoning, but rather that I am confused by the mind-altering properties of the CGTA gene and the chocolate soda. How do CGTA and chocolate soda influence peoples’ decisions? Do they alter peoples’ decision-types?
Maybe the CGTA gene gives you an itchy throat or makes you like to chew things. At any rate, chewing the gum is always the right choice (assuming the others costs of gum-chewing are negligible).
One intuition pump: if someone else forced you to chew gum, this wouldn’t have any bearing on whether you have CGTA, and it would lower your chances of abcess in either case, and so you’d be glad they’d done so. However, if someone else forced you to two-box, you’d be quite angry at having missed out on the million dollars.
In Newcomb’s problem, the result depends directly on your decision making process (by the definition of Omega/the Predictor), whereas with the gum example it doesn’t.
I think I see a difference in the intuitions surrounding the Newcomb’s and Solomon’s. This could explain why one-boxing and gum-chewing are compatible, and points to a strengthened version of Solomon’s that I think is a better parallel of Newcomb’s. I’ll try to explain it, but it’s too bad this is happening over the internet and not in front of a whiteboard.
tl;dr: Newcomb’s predictor is inescapable, but CGTA’s influence is escapable. Therefore you should one-box and chew gum. This is not an attempt at a new decision theory, just an argument that jogged my intuition.
Each problem involves a two-pronged causal processes emanating from a root cause. They state that because of the causal relationship between the root and the prongs, the prongs will likely “match”.
In Newcomb’s, the root cause is your mental state earlier in the day, and the prongs are your decision and the state of box B. The prongs “match” if the predictor predicted you correctly and filled box B accordingly.
According to the statement of the problem, the process that leads from the state of the root to matching prongs is inescapable. No matter how complex your decision process is, you cannot trick the predictor, because your decision process is what it is accurately predicting.
In Solomon’s, the root cause is your CGTA-status and the prongs are your gum-decision and your abscesses (or lack thereof). The prongs “match” if you are in the statistically common group for your decision (if you chew and have abscesses, or if you do not and do not).
Unlike Newcomb’s predictor, the process that makes your gum chewing match your throat abscesses seems escapable. The biological process that turns CGTA into throat abscesses is not predicting your decision process, so how could it make your throat-abscesses match your choice? The outcome in which you chew gum and don’t have abscesses seems very possible; after all, the people in the study didn’t know about any of this, did they? You should be able to act as though your decision is independent of your CGTA status, and take advantage of the benefits of gum-chewing.
Looking at the problems this way, I can see why I would one-box and chew gum. Newcomb’s predictor has been demonstrated to be inescapable in its accuracy, but CGTA hasn’t really been tested, and seems vulnerable to exploitation by well-considered decisions.
Consider an extension of Solomon’s problem, though, in which the link between gum-chewing and throat abscesses persists after the study. The link between CGTA and gum-chewing is so strong that, even after the study becomes well known, you can only convince CGTA-positive people to chew gum, and CGTA-negative people invariably decide not to, no matter the arguments. Well-known decision scientists publish papers arguing one way or another, and are always found to be CGTA-positive if they favor gum-chewing. Even after someone tests negative for CGTA, they refuse to chew gum, giving absurd-sounding reasons!
In this strengthened version of Solomon’s, I think that it now becomes reasonable to assume that CGTA is somehow deeply tied into human cognition, and attempting to escape its influence is as futile as trying to trick Newcomb’s Predictor.
… the chance that he could get a PhD just for this?
A Ph.D. in what? The subject matter fits into some odd interdisciplinary combination of Philosophy, Economics, Operations Research, AI/CompSci, and Statistics. In general, the research requirements for a PhD in CompSci are roughly equivalent to something like 4 published research papers plus a ~200 page dissertation containing material that can be worked into either a monograph or another half-dozen publishable papers. But there are other requirements besides research, and many institutions don’t like to allow people to “test out” of those requirements because it looks bad to the accrediting agencies.
Surprise at the quantity of work that had gone into it.
Alas, I totally failed to see the claimed “strange singularity at the heart of decision theory”.
My favourite bit was probably the speculations about agent boundaries—starting on p.108. Alas, from my POV, no mention of the wirehead problem.
Update 2011-06-26 regarding the new version. The bit that reads:
This manuscript was cut off here, but interested readers are suggested to look at these sources for more discussion:
...seems to have been deleted, and 3 pages worth of references have been added. The document seems to have had negligible additions, though—the bit on p.108 has moved back onto page 107. There seem to be a few more extra lines at the end about how “change” is a harmful concept in decision theory.
I expect people will be interested to hear that Eliezer’s TDT document has now been released for general consumption.
Does anyone else agree that, as a piece of expository writing, that document sucks bigtime?
111 pages! I got through about 25 and I was wondering why Eliezer thought I needed to hear about how his four friends had decided when presented with the Newcomb’s soda problem and how some people refer to this problem as Solomon’s problem. So, I decided to skim ahead until he started talking about TDT. So I skimmed and skimmed.
Finally, I got to section 14, entitled “The timeless decision procedure”. “Aha!”, I think. “Finally.” The first paragraph consists of a very long and confusing sentence which at least seems to deal with the timeless decision procedure.
It might be easier to understand if expressed as an equation or formula containing, you know, variables and things. So I read on, hoping to find something I can sink my teeth into. But then the second paragraph begins:
and closes with
As far as I can tell, the remainder of this section entitled “The timeless decision procedure” consists of this justification, though not from first principles, but rather using an example. And it doesn’t appear that Eliezer ever gets back to the task of providing a “formal presentation of a timeless decision algorithm”.
So, I skip forward to the end, hoping to read the conclusions. Instead I find:
Followed by a bibliography containing one entry—A chapter from a 1978 collection of articles on applications of decision theory.
″...was cut off here …”? Give me a break!
Let me know when you get it down to a dozen pages or so.
ETA: A cleaned up copy of the paper exists with a more complete bibliography and without the “manuscript was cut off here” closing.
I think this needs rewriting so it doesn’t sound so circular—and only mentions the word “conditional” once.
It seems to me that we can just say that it maximises utility—while maintaining an awareness that there may be other agents running its decision algorithm out there, in addition to all the other things it knows.
I think the stuff about “conditional upon the abstract computation returning that output” is pretty-much implied by the notion of utility maximisation.
Easier? That’s the opposite of true for this kind of material!
Easier if also expressed that way. You need the prose to know what the symbols mean, but the math itself is incredibly clearer when done as symbols.
I guess this is a case of “different strokes for different folks”. I will point out that it is fairly traditional for technical communication to contain formulas, equations, and/or pseudo-code. I believe the assumption behind this tradition is that such formal means of presentation are often clearer than expository text.
I am aware of the tradition. Yes, Eliezer’s piece does not include any semblance of technical rigour.
There is a reason the formal presentations include accompanying explanations. The mathematics for this kind of thing would be nigh incoherent and quite possibly longer than a verbal description. Expository text is utterly critical.
Incidentally, I have almost no doubt that “might be easier to understand” is not your real reason for demanding “you know, variables and things”. Some of your real reasons may actually be better in this instance.
Thanks for the link! I just read it all. The good: it’s very, very smooth reading—I know how well Eliezer can write, and even I was surprised at the quality—and it has some very lucid explanations of tricky matters (like why Pearlean causality is useful). The bad: it’s kinda rambling, contains many standard sci-fi LW arguments that feel out of place in a philosophy paper, and it doesn’t make any formal advances beyond what we already know here (I’d hoped to see at least one). The verdict: definitely read the first half if you’re confused about this whole “decision theory” controversy, it’ll get you unconfused in a pinch. Take the second half with a grain of salt because it’s still very raw (unmixed metaphor award!)
I wonder if it should be reformatted in LaTeX to pass item #1 from here.
It should be reformated in LaTeX so that it will look much much nicer.
I’m currently reading through the document, and yes, it definitely should. The present format is an unprofessional-looking eyesore, and the references are presented in a weird, clumsy, and inconsistent way. Using Latex/Bibtex would solve both problems easily and effectively.
(Personally, I can’t fathom why anyone capable of grasping the notion of a markup language would ever want to write a document longer than five pages in Word instead of Latex.)
7 and 8 are already a lost cause. :)
From a list of warning signs of a FAIL in an attempt to solve a famous problem:
The paper doesn’t build on (or in some cases even refer to) any previous work.
The paper wastes lots of space on standard material.
I would disagree that this paper doesn’t build on or take notice of previous work. It takes note of EDT and CDT and quite properly puts the focus on the point of departure of this work—specifically, the handling of contrafactuals. I’m quite happy with that aspect of the paper. My complaint was (8) that it wasted far too much space doing it. And, perhaps as a result of wasting so much time and space in preparation, it never reached its proper conclusion.
Also, it is not completely clear the Aronson’s list of warning signs really applies here. Eliezer is not solving a famous problem here. Most non-philosophers don’t think that a problem even exists. So, he does have to provide an explanation of why TDT is needed. Just not so much explanation.
Nor do I, and I would in any case suggest that some of them are screened off. There’s only so many times you can count ‘non-conventional’ as evidence.
I incidentally found some of the extra explanation handy purely as revision of various topics that it hadn’t particularly occurred to me were relevant.
I do hope someone goes ahead and finishes it. Including things like writing out that bibliography at the end and writing up the maths.
I must say I’m disappointed by the lack of rigor. On the other hand, I’m slightly relieved that he didn’t beat me to any of the stuff in the decision theory document I’m writing myself. So far, I have yet to see any formalization of decision theory that I would consider usable, other my own unfinished one.
I notice there seems to be an issue with the bibliography—there’s only one entry in it, but I’ve found at least one other citation in the text (Judea Pearl’s Causality cited on page 58) that’s not there. Are there any good collections of decision theory paper links out there?
If you have new formal arguments about decision theory, it would be much more useful to me (and others, I think) if you just posted them here in their current state instead. Or emailed them to the interested people.
Give a quick soundbite without context?
I’m approaching decision theory from from the perspective compilers approach optimizations: no approach is guaranteed to work always, but each one comes with a list of preconditions that you can check. I’m also summarizing some of the relevant work from compilers: automatic provably correct simplification, translation between forms, and a handy collection of normal forms to translate into.
For CDT, the precondition is a partial ordering over observation sets passed to the strategy such that the world program calls the strategy with observation sets only in increasing order, and there are finitely many possible observation sets. Then you can translate the program into continuation-passing style, and enumerate the possible invocations of the strategy function and their ordering. The last one in the order is guaranteed to have a continuation with no further invocations of the strategy function, which means you can try each possibility, simulate the results, and use that to determine the best answer. Then you can look at the second-to-last invocation, substitute the best answer to the last invocation into the continuation, and repeat; and so on for the set of all invocations to the strategy function. This works because you have a guarantee that when you compute your current position within the world-program and come up with a probability distribution over states to determine where you are, and then look at future continuations, changing result of any invocations of the strategy in those continuations does not affect the probability distribution over states.
I also have an example of a formalized decision-theory problem for which no optimal answer exists: name a number and that number is your utility. A corollary is that no decision theory can always give optimal answers, even given infinite computing power. This can be worked around by applying size bounds in various places.
I’m also drawing distinctions between strategies and decision theories (a strategy is an answer to one problem, a decision theory is an approach to generating strategies from problems); and between preference and utility (a preference is a partial order over outcomes; a utility function is a total order over outcomes where the outcomes are complete probability distributions, plus a linearity requirement).
So far, doesn’t sound good.
By that, do you mean that it sounds wrong, or that it sounds confused? If the former, I may need to reconsider; if the latter, I’m unsurprised because it’s much too short and doesn’t include any of the actual formalization. (That was not an excerpt from the draft I’m writing, but an attempt to summarize it briefly. I don’t think I did it justice.)
Doesn’t seem to address relevant questions or give interesting answers.
Ok, in that case I’m inclined to think that impression is just an artifact of how I summarized it, since my summary didn’t address the questions, but the longer paper I’m working on does, albeit only after building up proof and formalization techniques, which are the main focus.
Would something like UDT fit into your framework?
As far as I know, there are no cases where UDT suggests a decision and disagrees with mine. The differences are all in cases where UDT alone can’t be used to reach a decision.
I notice that the ideal causal diagram used in Part 2 (and based on pearls) is isomorphic to an example I use to teach CLIP, once you apply the substitution:
sprinkler on → a paperclip truck has over turned
rain → a clippy has haphazardly used up a lot of metal wire
sidewalk wet → paperclips are scattered across the ground
sidewalk slippery → many paperclips need to be moved to the safe zone
Thanks, Larks—how did you find out this was available? Is there a blog post or something somewhere? I didn’t see it on SIAI’s blog.
I’m glad to have this to read. I was surprised to find many examples and arguments that EY hadn’t given before (or at least formalized this way). I liked the Newcomb’s soda problem in particular. I had been worried that EY had presented enough of his TDT justification that someone could “scoop” him, but there’s a lot more depth to it. (Anyone up-to-date on the chance that he could get a PhD just for this?)
And I also appreciated that the modified the smoking lesion problem to be one where people aren’t distracted by their existing knowledge of smoking, and that this was the reason for transforming the example.
I read up to ~p. 35, and I think I have a better understanding now of the relevance of time consistency and how it varies across examples.
That said, I agree with the others who say it could use some mroe polish.
Yeah, Newcomb’s Soda and Solomon’s Problem are really interesting! If I faced the problems right now, I would one-box, skip the gum, and eat chocolate ice cream, because those choices put me in the groups that win, so I guess that classifies me as an evidentialist. At the same time, I haven’t reasoned out these conclusions thoroughly—I can’t argue formally against causal reasoning or dominance, or formally for evidentialism.
Looks like I have some more reading to do before I get this resolved.
Edit: All right, I think I can clarify my positions. I would not say that I am choosing based on evidential reasoning, but rather that I am confused by the mind-altering properties of the CGTA gene and the chocolate soda. How do CGTA and chocolate soda influence peoples’ decisions? Do they alter peoples’ decision-types?
Maybe the CGTA gene gives you an itchy throat or makes you like to chew things. At any rate, chewing the gum is always the right choice (assuming the others costs of gum-chewing are negligible).
Ah, maybe you can help me out. Why should I chew gum, but not two-box?
One intuition pump: if someone else forced you to chew gum, this wouldn’t have any bearing on whether you have CGTA, and it would lower your chances of abcess in either case, and so you’d be glad they’d done so. However, if someone else forced you to two-box, you’d be quite angry at having missed out on the million dollars.
In Newcomb’s problem, the result depends directly on your decision making process (by the definition of Omega/the Predictor), whereas with the gum example it doesn’t.
I think I see a difference in the intuitions surrounding the Newcomb’s and Solomon’s. This could explain why one-boxing and gum-chewing are compatible, and points to a strengthened version of Solomon’s that I think is a better parallel of Newcomb’s. I’ll try to explain it, but it’s too bad this is happening over the internet and not in front of a whiteboard.
tl;dr: Newcomb’s predictor is inescapable, but CGTA’s influence is escapable. Therefore you should one-box and chew gum. This is not an attempt at a new decision theory, just an argument that jogged my intuition.
Each problem involves a two-pronged causal processes emanating from a root cause. They state that because of the causal relationship between the root and the prongs, the prongs will likely “match”.
In Newcomb’s, the root cause is your mental state earlier in the day, and the prongs are your decision and the state of box B. The prongs “match” if the predictor predicted you correctly and filled box B accordingly.
According to the statement of the problem, the process that leads from the state of the root to matching prongs is inescapable. No matter how complex your decision process is, you cannot trick the predictor, because your decision process is what it is accurately predicting.
In Solomon’s, the root cause is your CGTA-status and the prongs are your gum-decision and your abscesses (or lack thereof). The prongs “match” if you are in the statistically common group for your decision (if you chew and have abscesses, or if you do not and do not).
Unlike Newcomb’s predictor, the process that makes your gum chewing match your throat abscesses seems escapable. The biological process that turns CGTA into throat abscesses is not predicting your decision process, so how could it make your throat-abscesses match your choice? The outcome in which you chew gum and don’t have abscesses seems very possible; after all, the people in the study didn’t know about any of this, did they? You should be able to act as though your decision is independent of your CGTA status, and take advantage of the benefits of gum-chewing.
Looking at the problems this way, I can see why I would one-box and chew gum. Newcomb’s predictor has been demonstrated to be inescapable in its accuracy, but CGTA hasn’t really been tested, and seems vulnerable to exploitation by well-considered decisions.
Consider an extension of Solomon’s problem, though, in which the link between gum-chewing and throat abscesses persists after the study. The link between CGTA and gum-chewing is so strong that, even after the study becomes well known, you can only convince CGTA-positive people to chew gum, and CGTA-negative people invariably decide not to, no matter the arguments. Well-known decision scientists publish papers arguing one way or another, and are always found to be CGTA-positive if they favor gum-chewing. Even after someone tests negative for CGTA, they refuse to chew gum, giving absurd-sounding reasons!
In this strengthened version of Solomon’s, I think that it now becomes reasonable to assume that CGTA is somehow deeply tied into human cognition, and attempting to escape its influence is as futile as trying to trick Newcomb’s Predictor.
A Ph.D. in what? The subject matter fits into some odd interdisciplinary combination of Philosophy, Economics, Operations Research, AI/CompSci, and Statistics. In general, the research requirements for a PhD in CompSci are roughly equivalent to something like 4 published research papers plus a ~200 page dissertation containing material that can be worked into either a monograph or another half-dozen publishable papers. But there are other requirements besides research, and many institutions don’t like to allow people to “test out” of those requirements because it looks bad to the accrediting agencies.
I scanned it. My initial reactions:
Surprise that the document existed;
TL;DR;
Surprise at the quantity of work that had gone into it.
Alas, I totally failed to see the claimed “strange singularity at the heart of decision theory”.
My favourite bit was probably the speculations about agent boundaries—starting on p.108. Alas, from my POV, no mention of the wirehead problem.
Update 2011-06-26 regarding the new version. The bit that reads:
...seems to have been deleted, and 3 pages worth of references have been added. The document seems to have had negligible additions, though—the bit on p.108 has moved back onto page 107. There seem to be a few more extra lines at the end about how “change” is a harmful concept in decision theory.