Two envelopes problem

> It would be boring and unsatisfying to accept a simple answer that explains how to get the "correct answer" without explaining what made it a paradox in the first place

Well, this is just a place where they do the math wrong. We can quibble about why, but it isn't a paradox. We both agree they do it wrong and we even have an overlap in the reason why - we both think they aren't comparing them relative to all the subgames they are in and we both think that this is required to reason correctly. So lets say we get to the correct relative EV of 0.

If both are worth the same, well, why not always switch? They are the same EV, right?

IMO this is the real paradox. We have a false equivocation. EV(policy, game) is not equal to EV(envelopes), because for the policy `ALWAYS_SWITCH` we get the logical contradiction that undefined = EV(envelope) or if you take the limit with a discount under a different formalism that 0 = EV(envelope)

When you tackle this problem starting there rather than at the other error the entire algorithm changes, because we need to modify the problem so that we can calculate EV with respect to policy. The algorithm changes enough so that it really annoys me, because I have this intuitive feeling that I'm changing the solution so much so that I'm no longer working within the spirit of the puzzle.

That is why I feel like "no true scottsman"; to use a metaphor, they gave me a coloring book and told me to color something beautiful but that I had to stay in the lines, but the lines of are a skunk and I want to draw a sunrise. I don't want to correct just the one particular step. I hate the framework that forces me into the confusion by falsely implying that if I get the right envelope EV, I know my EV.

> If both are worth the same, well, why not always switch?

I feel like this is a detour that we can avoid by adding a small cost to switching.

> I hate the framework that forces me into the confusion by falsely implying that if I get the right envelope EV, I know my EV.

I love it. I love it in the same way I love the trick of a magician who fools me. Also, the answer of "two EVs" is particularly satisfying for me, because it resembles a similar confusion in poker tournaments where you need to account for chip EV and dollar EV separately.

> I feel like this is a detour that we can avoid by adding a small cost to switching.

I don't think its a detour we actually want to avoid; the implications are really fascinating and help us to understand how to solve games more generally.

I agree with adding cost as a way to do it. Actually, that is why I've been saying it is zero or undefined. You know about dynamic programming right? Well, one of the reasons it was invented is kind of related to what we're talking about right now. There are these things called the bellman equations. They can look like this when you think of things in terms of a markov decision process.

https://wikimedia.org/api/rest_v1/media/math/render/svg/0e35...

Since they're recursively defined, you can compute them more quickly by caching the computation back to front then by computing them front to back. Anyways, putting aside that bit of trivia, do you see the symbol that looks a bit like a y? When we have an infinite sequence like we would if we kept switching you can set that term to be some constant, for example 0.9999999999999. That lets you take the limit, because the infinite sequence is obviously going to converge to zero.

Check out the formula again and notice one of the things that it does which the wikipedia article doesn't. Do you see the symbol for pi? That is talking about the concept of a policy function. A strategy, that way you compute what the agent would get if they played the game, rather than the envelope contents. Under this formalism we've doing an argmax over the policy function for the equation in order to get the policy that has the highest expected value.

> I love it. I love it in the same way I love the trick of a magician who fools me.

If you're actually interested in this sort of thing, I suggest checking out the poker research papers by Noam Brown. He had a talk at NIPS 2017 where he won best paper award. In 2019 he was runner up for best scientific advancement of the year. His work is about applying game simplification to poker to create a bluneprint of the game which is simpler, solving the blueprint using a modified form of counterfactual regret minimization, and then refining the solution during actual play with reach subgame solving. You might have heard of his work, because he was involved in the entire AI now better than humans at poker thing even in no limit breakthrough. I have some belief, but not certainty, that his paper contains a correction of someone who made this style of error. He mentions in the research paper that there was a mistake in another paper where they didn't account for the way subgames influence each other. I see that as the problem here, but I haven't read the other paper, so I don't know if it was the same category of error. Find the paper he referenced, read through it, and see if you're tricked. It is potentially a kind of real world two envelope problem to see if your tools for reasoning about these things is actually helping you to avoid the error in more complicated situations - though, since you disagree with me that this is about handling of imperfect information (or more precisely, the counterfactuals - there is a variant on wikipedia where they removed the probabilites, but counterfactual reasoning still applied to resolve the paradox) maybe you won't see it as related errors.

> I'm not a mathematician, so I don't understand expressions like "counterfactual part of subgames emerging" or "there is an identity implied by being under imperfect information". I appreciate you writing back at length, but the majority of what you wrote went straight over my head.

# Defining Counterfactuals

Consider a fair coin flip. You have {Heads, Tails}. Lets assume you are going to get heads, take it as a given - that is what actually happens. It actually happens. It is factual. However, for the purposes of analysis, sometimes it doesn't really matter that we know what happened. We need to consider all the cases that didn't happen. Tails in our analysis would be the counterfactual. These two situations, the factual heads and the counterfactual tails, they're associated with each other. There is a set {HT} that contains both of them. E.g P(head) = |{H}|/|{H, T}|.

When I'm saying counterfactual I'm referring to the events that we didn't assume to be factual, but which we want to keep track of. Probability kind of drops these terms when it says "assume" because it says |{H}|/|{H, T}| becomes |{H}|/|{H}| which is equal to one. This is perfectly fine from a math perspective. The thing is that just like saying 5/x = y lets us move to 5=xy is valid, it has some assumptions built into it. Namely that x can't be zero. Our counterfactuals are a bit like the x, because they offer an easily hidden constraint on what it is valid to do. If you reintroduce uncertainty about whether or not you are in H, you have to do so in a way that takes you back to having the set {H, T}.

Let me show you a practical example of that to make the point very clear:

Let the value of heads be one and the value of tails be zero. P(Head) = 0.5 But assume heads on the same coin flip. P(Head|assumptions) = 1.0 But assume heads again on the same coin flip. P(Head|assumptions) = 1.0. Now since P(Head) is actually true with 1/2 probability: Ev(Head) = 1/2*P(Head|assumption) + 1/2P(Head|assumption)?

Well, it is certainty true that 1/2P(Head|assumption) is a correct term. So you can't say this is wrong on the basis of the assumptions alone. Being very precise, the problem is the neglected counterfactual wasn't handled. Every term leading up to the equation was technically true, but obviously we just did something really weird right? And to be very precise, we neglected the counterfactual associated with heads.

# Defining subgames

Most games can be written out as a game tree. I showed one in my previous post. When you move down the tree, you are in a subgame of that tree.

In a perfect information game like chess if you are in a subgame, a portion lower on the tree, the other parts of the tree don't matter anymore. Your results are independent of the rest of the game tree.

In imperfect information though, just because you are in a subtree doesn't mean you know which subtree you are in. Your actual view into the game is through the information you have. Just like in the {HT} case you have to consider the potential that you have both H or T, in a subgame you have to consider the potential you are in every subgame that is reachable given the information you've seen.

Check out this to get a better sense of what I'm talking about: https://www.youtube.com/watch?v=EbKmZLp5HvA

> You know about dynamic programming right? Well, one of the reasons it was invented is kind of related to what we're talking about right now. There are these things called the bellman equations.

Yes, I've done dynamic programming in competitions. I'm not familiar with bellman equations, and the explanation you provided about convergence, policy functions, etc. went over my head, sorry.

> If you're actually interested in this sort of thing, I suggest checking out the poker research papers by Noam Brown [...] He mentions in the research paper that there was a mistake in another paper where they didn't account for the way subgames influence each other. I see that as the problem here, but I haven't read the other paper, so I don't know if it was the same category of error. Find the paper he referenced, read through it, and see if you're tricked. It is potentially a kind of real world two envelope problem to see if your tools for reasoning about these things is actually helping you to avoid the error in more complicated situations - though, since you disagree with me that this is about handling of imperfect information (or more precisely, the counterfactuals - there is a variant on wikipedia where they removed the probabilites, but counterfactual reasoning still applied to resolve the paradox) maybe you won't see it as related errors.

This sounds really interesting! It was a huge deal in the poker scene when the poker AI developed by Brown and Sandholm defeated pro human players. I read the first few pages of the paper now, but the paper becomes very math-heavy after that. I don't have the necessary background to understand the notation they use. That said, the "mistake" in the way "subgames influence each other" that you referenced, I suspect it was of a far simpler kind - the kind that Brown explains in chapter 2 which he concludes with:

This shows that a player’s optimal strategy in a subgame can depend on the strategies and outcomes in other parts of the game. Thus, one cannot solve a subgame using information about that subgame alone. This is the central challenge of imperfect-information games as opposed to perfect-information games.

Although Brown says "imperfect-information games" here, he actually means a specific type of imperfect-information games: the type where the opponent's strategy is not fixed. We're talking about games where your opponent can change their strategy in order to exploit weaknesses in your strategy. This property was a key requirement of the "coin toss" example game that he provided in chapter 2. If you modified the coin toss game such that the opponent's strategy was fixed, then the situation would change completely. Crucially, Two Envelopes Game is not one of those games where the "opponent" can adapt their strategy according to your strategy. The strategy of the opponent is fixed in Two Envelopes Game. That's why you can solve a single subgame in Two Envelopes game independently of other subgames, even though it's an imperfect-information game. If you are suspectful of this claim, we can verify it by simulations.