Today I finally came up with a simple example where TDT clearly loses and CDT clearly wins, and as a bonus, proves that TDT isn't reflectively consistent.

Omega comes to you and says

I'm hosting a game with 3 players. Two players are AIs I created running TDT but not capable of self-modification, one being a paperclip maximizer, the other being a staples maximizer. The last player is an AI you will design. When the game starts, my two AIs will first get the source code of your AI (which is only fair since you know the design of my AIs). Then 2 of the 3 players will be chosen randomly to play a one-shot true PD, without knowing who they are facing. What AI do you submit?

Say the payoffs of the PD are

  • 5/5 0/6
  • 6/0 1/1

Suppose you submit an AI running CDT. Then, Omega's AIs will reason as follows: "I have 1/2 chance of playing against a TDT, and 1/2 chance of playing against a CDT. If I play C, then my opponent will play C if it's a TDT, and D if it's a CDT, therefore my expected payoff is 5/2+0/2=2.5. If I play D, then my opponent will play D, so my payoff is 1. Therefore I should play C." Your AI then gets a payoff of 6, since it will play D.

Suppose you submit an AI running TDT instead. Then everyone will play C, so your AI will get a payoff of 5.

So you submit a CDT, whether you are running CDT or TDT. That's because explicitly giving the source code of your submitted AI to the other AIs makes the consequences of your decision the same under CDT and under TDT.

Suppose you have to play this game yourself instead of delegating it, you can self-modify, and the payoffs are large enough, you'd modify yourself from running TDT to running some other DT that plays D in this game! (Notice that I specified that Omega's AIs can't self-modify, so your decision to self-modify won't have the logical consequence that they also self-modify.)

It seems that I've given a counter-example to the claim that

the behavior of TDT corresponds to reflective consistency on a problem class in which your payoff is determined by the type of decision you make, but not sensitive to the exact algorithm you use apart from that

Or does my example fall outside of the specified problem class?

Wei_Dai wrote on 19 August 2009 07:08:23AM :

... Omega's AIs will reason as follows: "I have 1/2 chance of playing against a TDT, and 1/2 chance of playing against a CDT. If I play C, then my opponent will play C if it's a TDT, and D if it's a CDT ...

That seems to violate the secrecy assumptions of the Prisoner's Dilemma problem! I thought each prisoner has to commit to his action before learning what the other one did. What am I missing?

Thanks!

2Vladimir_Nesov11yThe problem is that "source code of your AI" is not a complete story, since your decisions as AI programmer also depended on the Omega AIs' code, and so what you give as the source of AI is already only one of the possible worlds that presupposes the behavior of Omega AIs.
14Eliezer Yudkowsky11yIf I wanted to defend the original thesis, I would say yes, because TDT doesn't cooperate or defect depending directly on your decision, but cooperates or defects depending on how your decision depends on its decision (which was one of the open problems I listed - the original TDT is for cases where Omega offers you straightforward dilemmas in which its behavior is just a direct transform of your behavior). So where one algorithm has one payoff matrix for defection or cooperation, the other algorithm gets a different payoff matrix for defection or cooperation, which breaks the "problem class" under which the original TDT is automatically reflectively consistent. Nonetheless it's certainly an interesting dilemma. Your comment here is actually pre-empting a comment that I'd planned to make after providing some of the background for the content of TDT. I'd thought about your dilemmas, and then did manage to translate into my terms a notion about how it might be possible to unilaterally defect in the Prisoner's Dilemma and predictably get away with it, provided you did so for unusual reasons. But the conditions on "unusual reasons" are much more difficult than your posts seem to imply. We can't all act on unusual reasons and end up doing the same thing, after all. How is it that these two TDT AIs got here, if not by act of Omega, if the sensible thing to do is always to submit a CDT AI? To introduce yet another complication: What if the TDTs that you're playing against, decide to defect unconditionally if you submit a CDT player, in order to give you an incentive to submit a TDT player? Given that your reason for submitting a CDT player involves your expectation about how the TDT players will respond, and that you can "get away with it"? It's the TDT's responses that make them "exploitable" by your decision to submit a CDT player - so what if they employ a different strategy instead? (This is another open problem - "who acts first" in timeless negotiations.) There m

Ingredients of Timeless Decision Theory

by Eliezer Yudkowsky 7 min read19th Aug 2009226 comments

49


Followup toNewcomb's Problem and Regret of Rationality, Towards a New Decision Theory

Wei Dai asked:

"Why didn't you mention earlier that your timeless decision theory mainly had to do with logical uncertainty? It would have saved people a lot of time trying to guess what you were talking about."

...

All right, fine, here's a fast summary of the most important ingredients that go into my "timeless decision theory".  This isn't so much an explanation of TDT, as a list of starting ideas that you could use to recreate TDT given sufficient background knowledge.  It seems to me that this sort of thing really takes a mini-book, but perhaps I shall be proven wrong.

The one-sentence version is:  Choose as though controlling the logical output of the abstract computation you implement, including the output of all other instantiations and simulations of that computation.

The three-sentence version is:  Factor your uncertainty over (impossible) possible worlds into a causal graph that includes nodes corresponding to the unknown outputs of known computations; condition on the known initial conditions of your decision computation to screen off factors influencing the decision-setup; compute the counterfactuals in your expected utility formula by surgery on the node representing the logical output of that computation.

To obtain the background knowledge if you don't already have it, the two main things you'd need to study are the classical debates over Newcomblike problems, and the Judea Pearl synthesis of causality.  Canonical sources would be "Paradoxes of Rationality and Cooperation" for Newcomblike problems and "Causality" for causality.

For those of you who don't condescend to buy physical books, Marion Ledwig's thesis on Newcomb's Problem is a good summary of the existing attempts at decision theories, evidential decision theory and causal decision theory.  You need to know that causal decision theories two-box on Newcomb's Problem (which loses) and that evidential decision theories refrain from smoking on the smoking lesion problem (which is even crazier).  You need to know that the expected utility formula is actually over a counterfactual on our actions, rather than an ordinary probability update on our actions.

I'm not sure what you'd use for online reading on causality.  Mainly you need to know:

  • That a causal graph factorizes a correlated probability distribution into a deterministic mechanism of chained functions plus a set of uncorrelated unknowns as background factors.
  • Standard ideas about "screening off" variables (D-separation).
  • The standard way of computing counterfactuals (through surgery on causal graphs).

It will be helpful to have the standard Less Wrong background of defining rationality in terms of processes that systematically discover truths or achieve preferred outcomes, rather than processes that sound reasonable; understanding that you are embedded within physics; understanding that your philosophical intutions are how some particular cognitive algorithm feels from inside; and so on.


The first lemma is that a factorized probability distribution which includes logical uncertainty - uncertainty about the unknown output of known computations - appears to need cause-like nodes corresponding to this uncertainty.

Suppose I have a calculator on Mars and a calculator on Venus.  Both calculators are set to compute 123 * 456.  Since you know their exact initial conditions - perhaps even their exact initial physical state - a standard reading of the causal graph would insist that any uncertainties we have about the output of the two calculators, should be uncorrelated.  (By standard D-separation; if you have observed all the ancestors of two nodes, but have not observed any common descendants, the two nodes should be independent.)  However, if I tell you that the calculator at Mars flashes "56,088" on its LED display screen, you will conclude that the Venus calculator's display is also flashing "56,088".  (And you will conclude this before any ray of light could communicate between the two events, too.)

If I was giving a long exposition I would go on about how if you have two envelopes originating on Earth and one goes to Mars and one goes to Venus, your conclusion about the one on Venus from observing the one on Mars does not of course indicate a faster-than-light physical event, but standard ideas about D-separation indicate that completely observing the initial state of the calculators ought to screen off any remaining uncertainty we have about their causal descendants so that the descendant nodes are uncorrelated, and the fact that they're still correlated indicates that there is a common unobserved factor, and this is our logical uncertainty about the result of the abstract computation.  I would also talk for a bit about how if there's a small random factor in the transistors, and we saw three calculators, and two showed 56,088 and one showed 56,086, we would probably treat these as likelihood messages going up from nodes descending from the "Platonic" node standing for the ideal result of the computation - in short, it looks like our uncertainty about the unknown logical results of known computations, really does behave like a standard causal node from which the physical results descend as child nodes.

But this is a short exposition, so you can fill in that sort of thing yourself, if you like.

Having realized that our causal graphs contain nodes corresponding to logical uncertainties / the ideal result of Platonic computations, we next construe the counterfactuals of our expected utility formula to be counterfactuals over the logical result of the abstract computation corresponding to the expected utility calculation, rather than counterfactuals over any particular physical node.

You treat your choice as determining the result of the logical computation, and hence all instantiations of that computation, and all instantiations of other computations dependent on that logical computation.

Formally you'd use a Godelian diagonal to write:

Argmax[A in Actions] in Sum[O in Outcomes](Utility(O)*P(this computation yields A []-> O|rest of universe))

(where P( X=x []-> Y | Z ) means computing the counterfactual on the factored causal graph P, that surgically setting node X to x, leads to Y, given Z)

Setting this up correctly (in accordance with standard constraints on causal graphs, like noncircularity) will solve (yield reflectively consistent, epistemically intuitive, systematically winning answers to) 95% of the Newcomblike problems in the literature I've seen, including Newcomb's Problem and other problems causing CDT to lose, the Smoking Lesion and other problems causing EDT to fail, Parfit's Hitchhiker which causes both CDT and EDT to lose, etc.

Note that this does not solve the remaining open problems in TDT (though Nesov and Dai may have solved one such problem with their updateless decision theory).  Also, although this theory goes into much more detail about how to compute its counterfactuals than classical CDT, there are still some visible incompletenesses when it comes to generating causal graphs that include the uncertain results of computations, computations dependent on other computations, computations uncertainly correlated to other computations, computations that reason abstractly about other computations without simulating them exactly, and so on.  On the other hand, CDT just has the entire counterfactual distribution rain down on the theory as mana from heaven (e.g. James Joyce, Foundations of Causal Decision Theory), so TDT is at least an improvement; and standard classical logic and standard causal graphs offer quite a lot of pre-existing structure here.  (In general, understanding the causal structure of reality is an AI-complete problem, and so in philosophical dilemmas the causal structure of the problem is implicitly given in the story description.)

Among the many other things I am skipping over:

  • Some actual examples of where CDT loses and TDT wins, EDT loses and TDT wins, both lose and TDT wins, what I mean by "setting up the causal graph correctly" and some potential pitfalls to avoid, etc.
  • A rather huge amount of reasoning which defines reflective consistency on a problem class; explains why reflective consistency is a rather strong desideratum for self-modifying AI; why the need to make "precommitments" is an expensive retreat to second-best and shows lack of reflective consistency; explains why it is desirable to win and get lots of money rather than just be "reasonable" (that is conform to pre-existing intuitions generated by a pre-existing algorithm); which notes that, considering the many pleas from people who want, but can't find any good intermediate stage between CDT and EDT, it's a fascinating little fact that if you were rewriting your own source code, you'd rewrite it to one-box on Newcomb's Problem and smoke on the smoking lesion problem...
  • ...and so, having given many considerations of desirability in a decision theory, shows that the behavior of TDT corresponds to reflective consistency on a problem class in which your payoff is determined by the type of decision you make, but not sensitive to the exact algorithm you use apart from that - that TDT is the compact way of computing this desirable behavior we have previously defined in terms of reflectively consistent systematic winning.
  • Showing that classical CDT, given self-modification ability, modifies into a crippled and inelegant form of TDT.
  • Using TDT to fix the non-naturalistic behavior of Pearl's version of classical causality in which we're supposed to pretend that our actions are divorced from the rest of the universe - the counterfactual surgery, written out Pearl's way, will actually give poor predictions for some problems (like someone who two-boxes on Newcomb's Problem and believes that box B has a base-rate probability of containing a million dollars, because the counterfactual surgery says that box B's contents have to be independent of the action).  TDT not only gives the correct prediction, but explains why the counterfactual surgery can have the form it does - if you condition on the initial state of the computation, this should screen off all the information you could get about outside things that affect your decision; then your actual output can be further determined only by the Godel-diagonal formula written out above, permitting the formula to contain a counterfactual surgery that assumes its own output, so that the formula does not need to infinitely recurse on calling itself.
  • An account of some brief ad-hoc experiments I performed on IRC to show that a majority of respondents exhibited a decision pattern best explained by TDT rather than EDT or CDT.
  • A rather huge amount of exposition of what TDT decision theory actually corresponds to in terms of philosophical intuitions, especially those about "free will".  For example, this is the theory I was using as hidden background when I wrote in "Causality and Moral Responsibility" that factors like education and upbringing can be thought of as determining which person makes a decision - that you rather than someone else makes a decision - but that the decision made by that particular person is up to you.  This corresponds to conditioning on the known initial state of the computation, and performing the counterfactual surgery over its output.  I've actually done a lot of this exposition on OBLW without explicitly mentioning TDT, like Timeless Control and Thou Art Physics for reconciling determinism with choice (actually effective choice requires determinism, but this confuses humans for reasons given in Possibility and Could-ness).  But if you read the other parts of the solution to "free will", and then furthermore explicitly formulate TDT, then this is what utterly, finally, completely, and without even a tiny trace of confusion or dissatisfaction or a sense of lingering questions, kills off entirely the question of "free will".
  • Some concluding chiding of those philosophers who blithely decided that the "rational" course of action systematically loses; that rationalists defect on the Prisoner's Dilemma and hence we need a separate concept of "social rationality"; that the "reasonable" thing to do is determined by consulting pre-existing intuitions of reasonableness, rather than first looking at which agents walk away with huge heaps of money and then working out how to do it systematically; people who take their intuitions about free will at face value; assuming that counterfactuals are fixed givens raining down from the sky rather than non-observable constructs which we can construe in whatever way generates a winning decision theory; et cetera.  And celebrating of the fact that rationalists can cooperate with each other, vote in elections, and do many other nice things that philosophers have claimed they can't.  And suggesting that perhaps next time one should extend "rationality" a bit more credit before sighing and nodding wisely about its limitations.
  • In conclusion, rational agents are not incapable of cooperation, rational agents are not constantly fighting their own source code, rational agents do not go around helplessly wishing they were less rational, and finally, rational agents win.

Those of you who've read the quantum mechanics sequence can extrapolate from past experience that I'm not bluffing.  But it's not clear to me that writing this book would be my best possible expenditure of the required time.

49