Kirby's Study Journal

[Kirby]

Science seems like a good thing. In contrast to having an idea that you just believe in, and then searching for facts to back it up, science says that there's this uncertainty. And then from this uncertainty, you maybe make up a theory. And you test it. Then you see what happens. Try to use some logic, and make a new experiment - ask a new question. See what reality tells you. This method is appealing, because it seems less prone to bias than other schools of thought. For example, religious or political dogma might tell you that "X is true". And then you can't question it. You just believe it. Maybe you have your doubts, but then you come up with ways to reconcile those doubts. Because after all, "X is true".

That's not science...

Then there's this idea that the scientific method is outdated. Before, we had limited data to work with, so we have to make this hypothesis first. Then do your test, see what happens, then find your results. Nowadays, data is abundant. Technology gives us tons of data to work with, so maybe instead of coming up with a hypothesis first, we just look at the data and have just the single question, "what do the patterns tell us?"...

I've been playing around with Elf, lately to evaluate games. It's a fun toy. I went through some pro games, and just a few minutes ago, I went through Andy Liu and Mark Lee's game from the Cotsen. Mark Lee won that game, and I reviewed it myself already. I already knew the result, but the game seemed pretty even, albeit complicated, to me the whole way through.

But when I reviewed with Elf, it says that Andy Liu had a solid lead up until the complicated fight in the middle. At that point, the tables turned... OK. From the other thread, I learned that these win percentages are kind of like Elf playing a few moves ahead and comparing resulting positions to things that it's been trained on already in its neural network. So somewhere, from the abyss, the evaluation there is possible which tells me that from thousands (millions?) of games played by Elf, we see that black wins more often.

Cool. I guess it's like the idea of having tons of data to work with. Then the patterns tell us: black is ahead. Kind of like a magic-8 ball when I think about it. I don't really know what face will show up when I shake the 8-ball, but at the end, I get an answer. But the difference is that the 8-ball is just random chance, whereas the answer I get from Elf is from patterns that have emerged from playing many, many games. In both cases, I just trust whatever answer is given to me.

Somehow, I feel like I've gone a bit in a circle. I'm back at the dogma that says, "X is true". Again, I can't question it. I just have to believe it. The one reconciling fact is that there is evidence that Elf wins a lot of games. So when Elf says "X is true", there's a good chance that Elf is right, due to its track record.

But it's still a little unsettling to me. I'm not into politics, so I have no idea if a politician with a good track record actually exists, but if s/he did, maybe this same logic would lead me to follow along with whatever that politician says when they say "X is true"...?

Probably not. I guess I have to come back to another basis of science, as I understand it: uncertainty. I said above that there's this uncertainty. You then make up a theory and test it. See how well it holds up.

When I'm reviewing games with Elf, I think I have to do the same thing. There's this uncertainty. Elf says black has 83% winrate. Then I should give that some degree of credit, given Elf's track record. But then I should test and explore variations there. I should try to understand it a little bit better. I can never be sure that black is ahead.

I can have my suspicions, and that's the best I can hope for...

I guess that's life.

I'm reminded of some medical issues my son was having a little while ago. We talk in terms of probabilities with the doctors sometimes. Some doctors have good credibility, but I have to remain suspicious. Maybe they are right, and maybe not. But again, I guess that's life.

In the meantime, I'd like to enjoy living... What does that mean here?

I suppose it means to enjoy the process of exploring variations. Seems a bit paradoxical, because in some sense, there's a satisfaction with things coming to an end. Knowing that, "yes, my group is alive" or "yes, I've won the game". Or maybe even knowing, "sadly, I've died" or "sadly, I've lost" give some sort of closure to a situation. But while the game is still going, and while the end hasn't come yet, that closure can't be there, yet.

For a given game, I suppose that gives the real-life result some meaning: there was some uncertainty with what was going on, but then closure came. Under different realities, maybe a different result could have occurred. But in that game, that was that.

Given the uncertainty in life... What's the story I want to create? I can't say I really know right now. Maybe that's why I spin my wheels writing posts like this on L19...

But what if that continues? I guess at some point in my life, there will be some sort of closure - health problem or otherwise. Was it a good story? I guess I can't really complain.

[EdLee]

Good evening, Kirby.

How were the train rides to and fro ? Did you sleep overnight on the train ? How many hours each way ? Would you recommend it ? Key: Fast Wifi ?!

[Kirby]

Good evening, Kirby.

How were the train rides to and fro ? Did you sleep overnight on the train ? How many hours each way ? Would you recommend it ? Key: Fast Wifi ?!

Hi, Ed. Do you mean to the Cotsen? I flew in. We are living in Michigan now, so taking a train might take awhile. I just took the weekend off for the two day tournament. I left Friday night and came back Monday morning and went to work. I was tired last Monday, but recovered by Tuesday.

[EdLee]

Hi Kirby, yes, Cotsen. The crazy rumors! I'd heard you took the train!

[dfan]

But when I reviewed with Elf, it says that Andy Liu had a solid lead up until the complicated fight in the middle. At that point, the tables turned... OK. From the other thread, I learned that these win percentages are kind of like Elf playing a few moves ahead and comparing resulting positions to things that it's been trained on already in its neural network. So somewhere, from the abyss, the evaluation there is possible which tells me that from thousands (millions?) of games played by Elf, we see that black wins more often.

Cool. I guess it's like the idea of having tons of data to work with. Then the patterns tell us: black is ahead. Kind of like a magic-8 ball when I think about it. I don't really know what face will show up when I shake the 8-ball, but at the end, I get an answer. But the difference is that the 8-ball is just random chance, whereas the answer I get from Elf is from patterns that have emerged from playing many, many games. In both cases, I just trust whatever answer is given to me.

Somehow, I feel like I've gone a bit in a circle. I'm back at the dogma that says, "X is true". Again, I can't question it. I just have to believe it. The one reconciling fact is that there is evidence that Elf wins a lot of games. So when Elf says "X is true", there's a good chance that Elf is right, due to its track record.

Yep, that's just how it is. The earliest versions of the network would have been just as happy to give you winrate numbers, and they would have been totally wrong. The only real way the network can "prove" its evaluations is for you to explore the variations with it and eventually come to the conclusion that it is indeed right. Sometimes (as will still happen sometimes with ladders, especially with ELF), you'll come to the conclusion that it's wrong.

You also don't have to "trust" or "believe" anything on faith, you can just use the data point "ELF thinks the game is 71% for Black" and do what you like with it, rather than the data point "the game is 71% for Black".

[Kirby]

Trust or faith comes in the conclusion you make from that data. If I read a sequence with my own brain, I have understanding of the process used to get there.

If I get a win rate of 71% for black, I trust elf's training data if I trust that black is ahead.

It's because my mind isn't a part of the training process, so I just get the answe like a black box.

Similar to if a pro comes over and says my move is bad, but doesn't give me a reason. I just have to trust him.

The sequences he shows me can help to some degree, but they are from his head and not mine.

[mhlepore]

Brian I'm guessing you are a two-boxer in Newcomb's problem.

[EdLee]

Newcomb's paradox

Hi mhlepore:

This is the first time I heard of this problem.
Could I ask you a question: the wikipedia page's description of "The Problem" states nothing about the accuracy or infallibility of the predictor:
is this intentionally left ambiguous as stated in the original formulation ?

The Wikipedia page mentions the infallibility of the predictor only later in the article. The article doesn't discuss whether the player knows anything about the infallibility of the predictor: so I assume the player doesn't.
Whether the predictor's accuracy is ~0%, ~100%, or exactly 50% ( quantum ): relevant or irrelevant ?

[Kirby]

Brian I'm guessing you are a two-boxer in Newcomb's problem.

I hadn't heard of this before, but with no information about the predictor's accuracy, choosing two-boxes seems to be the best choice (best payoff independent of the prediction). If there is a way to measure the historical accuracy of the predictor, then a different choice could be made.

It's not clear to me why this is a paradox.

[EdLee]

Hi Kirby,

It's not clear to me why this is a paradox.

Wikipedia: Whether the problem actually is a paradox is disputed.

From a purely practical standpoint: if the player is happy to take home at least $1000, then 2-boxes is the obvious choice. ( Collorary: if the player must take home at least $1,000 -- i.e. $0 is unacceptable for the player -- then 2-boxes is the only choice. )

However, I assume the point of the $1M is not the literal value, but let's change the situation such that the player needs the $1M for an operation to save a life ( or, we can easily change box B to be a life-or-death situation ), then it's completely not obvious 2-boxes is the "better" choice.
I'm guessing that's one area people argue over ?

My big question remains with the accuracy of the predictor.

[Kirby]

Hi Kirby,

From a purely practical standpoint: if the player is happy to take home at least $1000, then 2-boxes is the obvious choice. ( Collorary: if the player must take home at least $1,000 -- i.e. $0 is unacceptable for the player -- then 2-boxes is the only choice. )

However, I assume the point of the $1M is not the literal value, but let's change the situation such that the player needs the $1M for an operation to save a life ( or, we can easily change box B to be a life-or-death situation ), then it's completely not obvious 2-boxes is the "better" choice.
I'm guessing that's one area people argue over ?

Assuming the two-box choice, if you need $1,000,000, you still get $1,001,000 if the predictor predicted B. The *only* way I can see that it makes sense to pick a single box is if there is some reason to believe that the predictor is likely to be correct in his/her prediction. In that case, the optimal choice depends on how you quantify that belief (if you believe strongly that the predictor is correct, you might choose just one box).

But the problem doesn't seem to say anything about your belief in the predictor's prediction capability, so there's no reason to think the predictor is good at predicting...

Here's a simpler problem that seems similar, based on my understanding:
You can take this $100 bill that I have here, no questions asked. But FYI, the witch doctor down the street says that it's cursed, and will kill you.

Maybe if the witch doctor has a good track record, you might not take the $100... Better check his job history on LinkedIn...

[EdLee]

Kirby:

This clip says right away about the accuracy of the predictor -- which is different from Wikipedia; my big question remains: in the original formulation, does the predictor have 100% accuracy for eternity ?

Here.

[mhlepore]

Often Wikipedia is a good source, but seems here it is not.

As Newcomb's problem was explained to me long ago, the predictor has a very high historical accuracy rate (think > 95% but not perfect) at guessing whether someone will take both boxes or only the opaque box. If the predictor thinks you will take both boxes, he leaves nothing in the opaque box. If he thinks you will take only the opaque box, he leaves a million dollars in it for you.

* From an expected value perspective, one can make an argument that taking only the opaque box is the right move. More than 95% of the time, you will get the million dollars.

* From a dominant strategy perspective, one can argue for taking both boxes. The predictor has already decided, after all, to put the million in the opaque box or not, so whatever the predictor has done, I should take both boxes.

As for why it is sometimes referred to as a paradox, I suppose it depends on who you ask. Some argue that backward causality is an issue (my choosing both boxes is in effect causing the predictor to act in a certain way earlier in time), which is strange and could be labeled as paradoxical I suppose.

Anyway, put me down as a one-boxer. And once I get an AI Go program on my machine, I'll probably blindly take its advice.

[Uberdude]

Anyway, put me down as a one-boxer. And once I get an AI Go program on my machine, I'll probably blindly take its advice.

Blindly? beware of herp derp advice, here's Elf being dumb:

[Bill Spight]

Often Wikipedia is a good source, but seems here it is not.

As Newcomb's problem was explained to me long ago, the predictor has a very high historical accuracy rate (think > 95% but not perfect) at guessing whether someone will take both boxes or only the opaque box. If the predictor thinks you will take both boxes, he leaves nothing in the opaque box. If he thinks you will take only the opaque box, he leaves a million dollars in it for you.

* From an expected value perspective, one can make an argument that taking only the opaque box is the right move. More than 95% of the time, you will get the million dollars.

* From a dominant strategy perspective, one can argue for taking both boxes. The predictor has already decided, after all, to put the million in the opaque box or not, so whatever the predictor has done, I should take both boxes.

As for why it is sometimes referred to as a paradox, I suppose it depends on who you ask. Some argue that backward causality is an issue (my choosing both boxes is in effect causing the predictor to act in a certain way earlier in time), which is strange and could be labeled as paradoxical I suppose.

Anyway, put me down as a one-boxer. And once I get an AI Go program on my machine, I'll probably blindly take its advice.

IMO, the problem is not well defined. Backward causality is not a problem for physics in this case. However, that does not mean that the unknown past is determined. If my choice is not determined, then neither is what the predictor has done. Dominant strategies may not work with quantum mechanics if they depend upon reasoning by cases; with QM unknown cases can be in superposition.

OTOH, if we say that the choice is part of an iterated non-zero sum game, and that the predictor has a high prediction rate, we also need to know how the players who have taken the box have fared. Given all the relevant information, in that case the aura of paradox disappears.

Edit: OC, such an iterated non-zero sum game is not part of our experience. In fact, the "paradox" was constructed in such a way as to take us out of the world of experience. Given that, why should we think that there is a solution in the real world?

Edit 2: For a non-zero sum game that is closer to our everyday experience, consider the Prisoner's Dilemma. You can show by reasoning by cases that the dominant strategy is to defect (i.e., cooperate with the prosecutor). But if that is so, why do naive college freshmen, playing with strangers, cooperate and do better than defecting? You don't even have to iterate the game. The supposedly irrational cooperators win hands down. There is something wrong with reasoning by cases in non-zero sum games.