It seems like ICM is a **** approximation of your real tournament EV but someone decided to use that model and nobody challenged it.

Instead why doesnt someone make a program that observes MTTs and record millions of instances of chip stacks for each player at various points in the tournament, then record the actual finishing positions of those players once its over. Then feed it into some machine learning program that creates a model from real data?

ICM is not the only tournament equity model.

https://blog.gtowizard.com/when-does...icant-in-mtts/




It would be interesting to see some NN trained on real data though. That would make for an interesting project.

People are not stupid. Everybody knows that ICM is just one model. It seems to be a fairly good model which you can calculate pretty fast. People have already compared it to other models in (simplified) tournament simulations. In like the very endgame you could probably in theory run simulations (which respects all stack sizes, positions, exact blind timings, etc...) which would be more accurate but I don't know if people can run them fast. Your "idea" is too vague so I don't know if it would be any good (since I don't really know what exactly you want and how you want to use the data and what the purpose of the AI is). Also to beat a simple ICM model by adding AI would be no small task. What should the AI be handling? Also your idea would require a great deal of expertise in mathematics, programming, data analysis and AI which would come at a GREAT cost $$$. How could you justify such a cost? You want to run bots? Then there would be even more expenses and hurdles. If you can do all this you could probably make a lot more money legally in a different area. Also if someone actually did do something like this we wouldn't know it (though I doubt it for the reason above).

How can you say this? Since nobody knows the true EVs of stacks you can’t compare ICM to the reality and can only guess at how good it is. My suggestion would give an answer based on the reality

Also it’s not fast at all - it’s so slow you can’t calculate beyond 12 players or something


I’ve only done casual programming but wouldnt be hard to do this if wanted to. The only part I’d do suboptimally is fitting the best model to the data. Not going to though since am a cash game player predominately and was only throwing the idea out there

Considering how there’s bigger and more high rollers than ever you could probably get around $1m from a good solution if you sold it to the right people

As I said people have done simplified tournament simulations where the players played different models and ICM was performing good. And you can calculate it pretty fast (you don't calculate it exactly, you have some algo which converges to it with good speed). Also you can judge the model by itself by the assumptions it makes.

If someone who casually programs could do it something better has been done already. As said your idea as presented is very vague, it is not clear to me what exactly you are trying to do.

Also I doubt anyone would pay 1mio$...

In my previous career back in the 1980s, mathematical modeling was something I was involved in. And a good mathematical model is usually an equation that is supposed to be simple yet still does a decent job of answering a lot of otherwise difficult questions. And I think that's what ICM does.

So, can a more accurate way be developed (or perhaps already exists) than ICM? Probably. But will this method be simpler? Probably not. And again, a good mathematical model is something that should be simple.

Perhaps the most famous example of a simple mathematical model that answered complicated questions was when a man named Copernicus said the planets orbited the sun in circles. Today, we know that the planets actually orbit the sun in complex elliptical orbits. But the simple model based on circles, even though it took years to be accepted, did answer a lot of questions, and I think that's the way to view ICM.

For those interested, my original writing on ICM can be found in the chapter "To Rebuy or Not to Rebuy" starting on page 243 in the expanded edition of my Gambling Theory book: https://www.amazon.com/Gambling-Theo...s%2C166&sr=1-2

Mason

Yeah, I'm not a tournament player but as one you probably just want to see the shift in strategy compared to chipEV and for that ICM should be good enough. More precise models would probably only really be useful for bots/rta play IMO. But then again if someone is cheating they can even collude in non obvious ways etc... which would most likely trump any EV gain of a better model by a lot.

Could you explain why you think a simple model is better than a more accurate one? Whether you use ICM or a hypothetical high-accuracy model, you plug the parameters into a computer and it's a "black box" that spits out an answer (I know you could calculate ICM by hand if you wanted but nobody does that).

I don't see why the complexity of what happens inside the black-box matters, but perhaps I'm missing something

Are you talking about bots/rta playing or legit poker? For legit poker you, as said, probably just want to see the shift in play and for that ICM should be good enough (to see the general direction) (I'm not sure to what accuracy tournament players memorize spots but I doubt they learn hundred of spots by heart). For bots/rta play obviously the more accurate model (if it runs fast) is preferred but as said if you are cheating there is much more EV gained elsewhere so this should be so minute that it shouldn't really matter that much. Also the biggest discrepancies will probably be in the end game and there is no way to learn enough end game scenarios by heart to make it useful (think about how many combinations of stacks, blind levels and payouts there are).

The underlying assumption of ICM theory is that players are all of equal capability. That, of course, is almost certainly not true. In the following blog post, I suggest one way to account for varying capability by using a player’s ROI data to develop what I call a Capability Factor. How useful this is I’ll leave to others


http://www.tumblr.com/holdemmatholog...ual-capability

idk why you're so interested in shooting down literally everything I say.

But if you don't see how being able to use more accurate solutions to study is valuable, there's no point discussing with you. For example every good mtt reg prefers to use HRC with FGS than ICM when possible.

And ICM doesn't show you the correct shifts in play. Because for instance it doesn't consider at all the extra future EV of having a big stack. And regs do want to know exactly how spots play in equilibrium, not only see shifts

Nice, thanks. That's another thing I'd never really considered before

Simpler models are easier to use and understand. If that’s not the case, then a more complex model that gives a more accurate answer would be better. However, if a more complex model, even if all it means to get an answer is a little more computer input, is not understood by a lot of players and poker management, then that's probably not a good way to go.

Mason

That's true for the vanilla algorithm, but you can get very very close numerical approximations of ICM even for huge fields. We do use these in HRC and they scale fine even for thousands of players.

If you mean simulated play then yeah, done that. And you can certainly get some slight improvements over ICM with variants that are fitted to specific game types or structures. But have not been at the point where I felt that the introduction of a new specific model would really be justified over just using the general "simple" one.

ICM is a pretty good model for approximating adjustments that should be made based on tournament payout structures, especially on the money bubble and final table when payout implications matter most. The idea of a "bubble factor," which is some additional amount of equity over chip EV that is required to risk your chips is an especially useful way to think about these spots.

IMO it's most useful because we can look at various spots to learn patterns. The simplicity helps humans develop a practical intuition about how much we should be adjusting in various spots that's simple enough we can use it at the tables.

That being said I agree with the OP that there's an opportunity to use machine learning to develop a more accurate system. In the past I thought that you could plug in massive amounts of data from online tournaments and train the AI neural network on real data. However the challenge is that there are just too many variables that need to be accounted for to come up with a program that can provide meaningful solutions for humans.

For example you're going to have some examples of recreational players making huge blunders, and that would end up in the aggregate data. The resulting solution would then be more akin to an exploitative solution. If player tendencies changed the solution would also change.

It could be useful to know the best strategy against the pool, but as we all know you're going to want to play very differently versus a recreational player compared to playing against a top pro.

The other problem with too many variables is that they all need to be accounted for somehow. Different payout structures, stack size distributions, blind structures and many more variables all effect the result. How do you code the AI to account for all these things in a way that will provide meaningful practical solutions that humans can use?

It's no small task. Anyway this is an interesting discussion.