Two-boxers think that decisions are things that can just fall out of the sky uncaused. (This can be made precise by a suitable description of how two-boxers set up the relevant causal diagram; I found Anna Salamon's explanation of this particularly clear.) This is a view of how decisions work driven by intuitions that should be dispelled by sufficient knowledge of cognitive and / or computer science. I think acquiring such background will make you more sympathetic to the perspective that one should think in terms of winning agent types and not winning decisions.
I also think there's a tendency among two-boxers not to take the stakes of Newcomb's problem seriously enough. Suppose that instead of offering you a million dollars Omega offers to spare your daughter's life. Now what do you do?
Basically: EDT/UDT has simple arguments in its favor and seems to perform well. There don't seem to be any serious arguments in favor of CDT, and the human intuition in its favor seems quite debunkable. So it seems like the burden of proof is on CDT, to justify why it isn't crazy. It may be that CDT has met that burden, but I'm not aware of it.
A. The dominance arguments in favor of two-boxing seem quite weak. They tend to apply verbatum to playing prisoner's dilemmas against a mirror (If the mirror cooperates you'd prefer defect, if the mirror defects you'd prefer defect, so regardless of the state of nature you'd prefer defect). So why do you not accept the dominance argument for a mirror, but accept it in the case of Newcomb-like problems? To discriminate the cases it seems you need to make an assumption of no causal connection, or a special role for time, in your argument.
This begs the question terribly---why is a causal connection privileged? Why is the role of time privileged? As far as I can tell these two things are pretty arbitrary and unimportant. I'm not aware of any strong philosophical arguments for CDT, besides "it seems intuitively sensible to a human," and...
I'm happy to learn that you consider UDT a variant of EDT, because after thinking about these issues for awhile my current point of view is that some form of EDT is obviously the correct thing to do, but in standard examples of EDT failing the relevant Bayesian updates are being performed incorrectly. The problem is that forcing yourself into a reference class by performing an action doesn't make it reasonable for you to reason as if you were a random sample from that reference class, because you aren't: you introduced a selection bias. Does this agree with your thoughts?
The intuition pump that got me to be a very confident one-boxer is the idea of submitting computer code that makes a decision, rather than just making a decision.
In this version, you don't need an Omega - you just need to run the program. It's a lot more obvious that you ought to submit a program that one-boxes than it is obvious that you ought to one-box. You can even justify this choice on causal decision-theory grounds.
With the full Newcomb problem, the causality is a little weird. Just think of yourself as a computer program with partial self-awareness. Deciding whether to one-box or two-box updates the "what kind of decision-making agent am I" node, which also caused Omega to either fill or not fill the opaque box.
Yes, it's wonky causality - usually the future doesn't seem to affect the past. Omega is just so unlikely that given that you're talking to Omega, you can justify all sorts of slightly less unlikely things.
So I'm turning to you guys to ask for help figuring out whether I should be a staunch one-boxer too.
Only if you like money.
The optimal thing would be to have Omega think that you will one-box, but you actually two box. You'd love to play Omega for a fool, but the problem explicitly tells you that you can't, and that Omega can somehow predict you.
Omega has extremely good predictions. if you've set your algorithm in such a state that Omega will predict that you one-box, you will be unable to do anything but one-box - your neurons are set in place, causal lines have already insured your decision, and free will doesn't exist in the sense that you can change your decision after the fact.
The problem is no free lunch. Any decision theory is going to fail somewhere. The case for privileging Newcomb as a success goal over all other considerations has not, in fact, been made.
So I raised this problem too, and I got a convincing answer to it. The way I raised it was to say that it isn't fair to fault CDT for failing to maximise expected returns in Newcomb's problem, because Newcomb's problem was designed to defeat CDT and we can design a problem to defeat any decision theory. So that can't be a standard.
The response I got (at least, my interpretation of it) was this: It is of course possible to construct a problem in which any decision theory is defeated, but not all such problems are equal. We can distinguish in principle between problems that can defeat any decision procedure (such as 'omega gives you an extra million for not using X', where X is the decision procedure you wish to defeat) and problems which defeat certain decision procedures but cannot be constructed so as to defeat others. Call the former type 1 problems, and the latter type 2 problems. Newcomb's problem is a type 2 problem, as is the prisoner's dilemma against a known psychological twin. Both defeat CDT, but not TDT, and cannot be constructed so as to defeat TDT without becoming type 1. TDT is aimed (though I think not yet successful) at being able to solve all type 2 problems.
So if ...
[Saying same thing as everyone else, just different words. Might work better, might not.]
Suppose once Omega explains everything to you, you think 'now either the million dollars are there or aren't and my decision doesn't affect shit.' True, your decision now doesn't affect it - but your 'source code' (neural wiring) contains the information 'will in this situation think thoughts that support two-boxing and accept them.' So, choosing to one-box is the same as being the type of agent who'll one-box.
The distinction between agent type and decision is artifici...
My problem with causal decision theory is that it treats the past different from the future for no good reason. If you read the quantum physics sequence, particularly the part about timeless physics, you will find that time is most likely not even an explicit dimension. The past is more likely to be known, but it's not fundamentally different from the future.
The probability of box A having money in it is significantly higher given that you one box then the probability given that you do not. What more do you need to know?
Here is another way to think about this problem.
Imagine if instead of Omega you were on a futuristic game show. As you go onto the show, you enter a future-science brain scanner that scans your brain. After scanning, the game show hosts secretly put the money into the various boxes behind stage.
You now get up on stage and choose whether to one or two box.
Keep in mind that before you got up on the show, 100 other contestants played the game that day. All of the two-boxers ended up with less money than the one-boxers. As an avid watcher of the show, you c...
Didn't we have a thread about this really recently?
Anyhow, to crib from the previous thread - an important point is reflective equilibrium. I shouldn't be able to predict that I'll do badly - if I know that, and the problem is "fair" in that it's a decision-determined, I can just make the other decision. Or if I'm doing things a particular way, and I know that another way of doing things would be better, and the problem is "fair" in that I can choose how to do things, I can just do things the better way. To sit and stew and lose anyho...
Okay, those with a two-boxing agent type don't win but the two-boxer isn't talking about agent types. They're talking about decisions. So they are interested in what aspects of the agent's winning can be attributed to their decision and they say that we can attribute the agent's winning to their decision if this is caused by their decision. This strikes me as quite a reasonable way to apportion the credit for various parts of the winning.
Do I understand it correctly that you're trying to evaluate the merits of a decision (to two-box) in isolation of the decision procedure that produced it? Because that's simply incoherent if the payoffs of the decision depend on your decision procedure.
To put it succinctly, Omega knows me far better than I know myself. I'm not going to second guess him/her.
The case of CDT vs Newcomb-like problems to me has a lot of similarity with the different approaches to probability theory.
In CDT you are considering only one type of information, i.e. causal dependency, to construct decision trees. This is akin to define probability as the frequency of some process, so that probability relations become causal ones.
Other approaches like TDT construct decisions using causal and logical dependency, as the induction logic approach to probability does.
The Newcomb is not designed to be "unfair" to CDT, it is designed...
Okay, those with a two-boxing agent type don't win but the two-boxer isn't talking about agent types. They're talking about decisions.
The problem doesn't care whether you are the type of agent who talks about agent types or the type of agent who talks about decisions. The problem only cares about which actions you choose.
Consider the following two mechanisms for a Newcomb-like problem.
A. T-Omega offers you the one or two box choice. You know that T-Omega used a time machine to see if you picked one or two boxes, and used that information to place/not place the million dollars.
B. C-Omega offers you the one or two box choice. You know that C-Omega is con man, that pretends great predictive powers on each planet he visits. Usually he fails, but on Earth he gets lucky. C-Omega uses a coin flip to place/not place the million dollars.
I claim the correct choice is to one box T...
[Two boxers] are interested in what aspects of the agent's winning can be attributed to their decision and they say that we can attribute the agent's winning to their decision if this is caused by their decision. This strikes me as quite a reasonable way to apportion the credit for various parts of the winning.
What do you mean by "the agent's winning can be attributed to their decision"? The agent isn't winning! Calling losing winning strikes me as a very unreasonable way to apportion credit for winning.
It would be helpful to me if you defined...
The one problem I had with Yudkowsky's TDT paper (which I didn't read very attentively, mind you, so correct me if I'm wrong) was the part where he staged a dramatic encounter where a one-boxer was pleading with a wistful two-boxing agent who wished he was a one-boxer to change his algorithm to choose just one box. It occurred to me that even if the two-boxer listened to her, then his algorithm would have been altered by totally external factors. For the superintelligence setting up the problem to have predicted his change of mind, he would have had to sim...
I have sympathy with both one-boxers and two-boxers in Newcomb's problem. Contrary to this, however, many people on Less Wrong seem to be staunch and confident one-boxers. So I'm turning to you guys to ask for help figuring out whether I should be a staunch one-boxer too. Below is an imaginary dialogue setting out my understanding of the arguments normally advanced on LW for one-boxing and I was hoping to get help filling in the details and extending this argument so that I (and anyone else who is uncertain about the issue) can develop an understanding of the strongest arguments for one-boxing.