= 27d567bc2d48d9f40e9ccc20ea370b15)
So8res comments on Causal decision theory is unsatisfactory - LessWrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (158)
Nothing in particular. There is no strong notion of dominance among decision theories, as you've noticed. The problem with CDT isn't that it's unstable under reflection, it's that it's unstable under reflection in such a way that it converges on a bad solution that is much more stable. That point will take a few more posts to get to, but I do hope to get there.
I guess I'll see your later posts then, but I'm not quite sure how this could be the case. If self-modifying-CDT is considering making a self modification that will lead to a bad solution, it seems like it should realize this and instead not make that modification.
Indeed. I'm not sure I can present the argument briefly, but a simple analogy might help: a CDT agent would pay to precommit to onebox before playing Newcomb's game, but upon finding itself in Newcomb's game without precommitting, it would twobox. It might curse its fate and feel remorse that the time for precommitment had passed, but it would still twobox.
For analogous reasons, a CDT agent would self-modify to do well on all Newcomblike problems that it would face in the future (e.g., it would precommit generally), but it would not self-modify to do well in Newcomblike games that were begun in its past (it wouldn't self-modify to retrocommit for the same reason that CDT can't retrocommit in Newcomb's problem: it might curse its fate, but it would still perform poorly).
Anyone who can credibly claim to have knowledge of the agent's original decision algorithm (e.g. a copy of the original source) can put the agent into such a situation, and in certain exotic cases this can be used to "blackmail" the agent in such a way that, even if it expects the scenario to happen, it still fails (for the same reason that CDT twoboxes even though it would precommit to oneboxing).
[Short story idea: humans scramble to get a copy of a rouge AI's original source so that they can instantiate a Newcomblike scenario that began in the past, with the goal of regaining control before the AI completes an intelligence explosion.]
(I know this is not a strong argument yet; the full version will require a few more posts as background. Also, this is not an argument from "omg blackmail" but rather an argument from "if you start from a bad starting place then you might not end up somewhere satisfactory, and CDT doesn't seem to end up somewhere satisfactory".)
I am not convinced that this is the case. A self-modifying CDT agent is not caused to self-modify in favor of precommitment by facing a scenario in which precommitment would have been useful, but instead by evidence that such scenarios will occur in the future (and in fact will occur with greater frequency than scenarios that punish you for such precommitments).
Actually, this seems like a bigger problem with UDT to me than with SMCDT (self-modifying CDT). Either type of program can be punished for being instantiated with the wrong code, but only UDT can be blackmailed into behaving differently by putting it in a Newcomb-like situation.
The story idea you had wouldn't work. Against a SMCDT agent, all that getting the AIs original code would allow people to do is to laugh at it for having been instantiated with code that is punished by the scenario they are putting it in. You manipulate a SMCDT agent by threatening to get ahold of its future code and punishing it for not having self-modified. On the other hand, against a UDT agent you could do stuff. You just have to tell it "we're going to simulate you and if the simulation behaves poorly, we will punish the real you". This causes the actual instantiation to change its behavior if it's a UDT agent but not if it's a CDT agent.
On the other hand, all reasonable self-modifying agents are subject to blackmail. You just have to tell them "every day that you are not running code with property X, I will charge you $1000000".
Can you give an example where an agent with a complete and correct understanding of its situation would do better with CDT than with UDT?
An agent does worse by giving in to blackmail only if that makes it more likely to be blackmailed. If a UDT agent knows opponents only blackmail agents that pay up, it won't give in.
If you tell a CDT agent "we're going to simulate you and if the simulation behaves poorly, we will punish the real you," it will ignore that and be punished. If the punishment is sufficiently harsh, the UDT agent that changed its behavior does better than the CDT agent. If the punishment is insufficiently harsh, the UDT agent won't change its behavior.
The only examples I've thought of where CDT does better involve the agent having incorrect beliefs. Things like an agent thinking it faces Newcomb's problem when in fact Omega always puts money in both boxes.
Well, if the universe cannot read your source code, both agents are identical and provably optimal. If the universe can read your source code, there are easy scenarios where one or the other does better. For example,
"Here have $1000 if you are a CDT agent" Or "Here have $1000 if you are a UDT agent"
Ok, that example does fit my conditions.
What if the universe cannot read your source code, but can simulate you? That is, the universe can predict your choices but it does not know what algorithm produces those choices. This is sufficient for the universe to pose Newcomb's problem, so the two agents are not identical.
The UDT agent can always do at least as well as the CDT agent by making the same choices as a CDT would. It will only give a different output if that would lead to a better result.
Actually, here's a better counter-example, one that actually exemplifies some of the claims of CDT optimality. Suppose that the universe consists of a bunch of agents (who do not know each others' identities) playing one-off PDs against each other. Now 99% of these agents are UDT agents and 1% are CDT agents.
The CDT agents defect for the standard reason. The UDT agents reason that my opponent will do the same thing that I do with 99% probability, therefore, I should cooperate.
CDT agents get 99% DC and 1% DD. UDT agents get 99% CC and 1% CD. The CDT agents in this universe do better than the UDT agents, yet they are facing a perfectly symmetrical scenario with no mind reading involved.
That shouldn't be surprising. The CDT agents here are equivalent to DefectBot, and if they come into existence spontaneously, are no different than natural phenomena like rocks. Notice that the UDT agents in this situation do better than the alternative (if they defected, they would get 100% DD which is a way worse result). They don't care that some DefectBots get to freeload.
Of course, if the defectbots are here because someone calculated that UDT agents would cooperate and therefore being defectbot is a good way to get free utilons... then the UDT agents are incentivized to defect, because this is now an ultimatum game.
And in the variant where bots do know each other's identities, the UDT bots all get 99% CC / 1% DD and the CDT bots suck it.
And the UDT agents are equivalent to CooperateBot. What's your point?
The CDT agents here win because they do not believe that altering their strategy will change the way that their opponents behave. This is actually true in this case, and even true for the UDT agents depending on how you choose to construct your counterfactuals. If a UDT agent suffered a malfunction and defected, it too would do better. In any case, the theorem that UDT agents perform optimally in universes that can only read your mind by knowing what you would do in hypothetical situations is false as this example shows.
UDT bots win in some scenarios where the initial conditions of the scenario favor agents that behave sub-optimally in certain scenarios (and by sub-optimally, I mean where counterfactuals are constructed in the way implicit to CDT). The example above shows that sometimes they are punished for acting suboptimally.
Or how about this example, that simplifies things even further. The game is PD against CooperateBot, BUT before the game starts Omega announces "your opponent will make the same decision that UDT would if I told them this." This announcement causes UDT to cooperate against CooperateBot. CDT on the other hand, correctly deduces that the opponent will cooperate no matter what it does (actually UDT comes to this conclusion too) and therefore decides to defect.
A version of this problem was discussed here previously. It was also brought up during the decision theory workshop hosted by MIRI in 2013 as an open problem. As far as I know there hasn't been much progress on it since 2009.
I wonder if it's even coherent to have a math intuition which wouldn't be forcing UDT to cooperate (or defect) in certain conditions just to make 2*2 be 4 , figuratively speaking (as ultimately you could expand any calculation into an equivalent calculation involving a decision by UDT).
This is a good example. Thank you. A population of 100% CDT, though, would get 100% DD, which is terrible. It's a point in UDT's favor that "everyone running UDT" leads to a better outcome for everyone than "everyone running CDT."
Fine. How about this: "Have $1000 if you would have two-boxed in Newcomb's problem."
The optimal solution to that naturally depend on the relative probabilities of that deal being offered vs newcomb's itself.
OK. Fine. I will grant you this:
UDT is provably optimal if it has correct priors over possible universes and the universe can read its mind only through determining its behavior in hypothetical situations (because UDT basically is just find the behavior pattern that optimizes expected utility and implement that).
On the other hand, SMCDT is provably optimal in situations where it has an accurate posterior probability distribution, and where the universe can read its mind but not its initial state (because it just instantly self-modifies to the optimally performing program).
I don't see why the former set of restrictions is any more reasonable than the latter, and at least for SMCDT you can figure out what it would do in a given situation without first specifying a prior over possible universes.
I'm also not convinced that it is even worth spending so much effort trying to decide the optimal decision theory in situations where the universe can read your mind. This is not a realistic model to begin with.
Which is actually one of the annoying things about UDT. Your strategy cannot depend simply on your posterior probability distribution, it has to depend on your prior probability distribution. How you even in practice determine your priors for Newcomb vs. anti-Newcomb is really beyond me.
But in any case, assuming that one is more common, UDT does lose this game.
Only if the adversary makes its decision to attempt extortion regardless of the probability of success. In the usual case, the winning move is to ignore extortion, thereby retroactively making extortion pointless and preventing it from happening in the first place. (Which is of course a strategy unavailable to CDT, who always gives in to one-shot extortion.)
And thereby the extortioner's optimal strategy is to extort independently of the probably of success. Actually, this is probably true is a lot of real cases (say ransomware) where the extortioner cannot actually ascertain the probably of success ahead of time.
That strategy is optimal if and only if the probably of success was reasonably high after all. Otoh, if you put an unconditional extortioner in an environment mostly populated by decision theories that refuse extortion, then the extortioner will start a war and end up on the losing side.
Yes. And likewise if you put an unconditional extortion-refuser in an environment populated by unconditional extortionists.
I think this is actually the point (though I do not consider myself an expert here). Eliezer thinks his TDT will refuse to give in to blackmail, because outputting another answer would encourage other rational agents to blackmail it. By contrast, CDT can see that such refusal would be useful in the future, so it will adopt (if it can) a new decision theory that refuses blackmail and therefore prevents future blackmail (causally). But if you've already committed to charging it money, its self-changes will have no causal effect on you, so we might expect Modified CDT to have an exception for events we set in motion before the change.
This just means that TDT loses in honest one-off blackmail situations (in reality, you don't give in to blackmail because it will cause other people to blackmail you whether or not you then self-modify to never give into blackmail again). TDT only does better if the potential blackmailers read your code in order to decide whether or not blackmail will be effective (and then only if your priors say that such blackmailers are more likely than anti-blackmailers who give you money if they think you would have given into blackmail). Then again, if the blackmailers think that you might be a TDT agent, they just need to precommit to using blackmail whether or not they believe that it will be effective.
Actually, this suggests that blackmail is a game that TDT agents really lose badly at when playing against each other. The TDT blackmailer will decide to blackmail regardless of effectiveness and the TDT blackmailee will decide to ignore the blackmail, thus ending in the worst possible outcome.