"What decision do I make now?" : Math will not get you there

[Cyrus]

I'm reposting here the arguments made by Utah, in another forum of this website and on an unrelated subject originally. I'm not sure what Utah is trying to say exactly except that he -and his poker team- are after optimal play in each and every situation, irrespective of past results. And aren't we all? But I wish Utah would elaborate a bit, because which data exactly matter in his analysis and which data don't is not too clear.

Here is Utah, in own words, paraphrased only in one or two points in order to eliminate some irrelevant topical matters from the original thread.

I don't know how to measure results objectively. How does one do that? In reality, it is next to impossible to truly do [that] with anything of any complexity.

[Poker Tracker and StatKing] do nothing for the player and they will not help you improve your game. I created a much better tool that takes a completely different approach.

Let's take a simple stat - preflop aggressiveness. Let's say that I raised preflop 20% of the time. Is that good or bad? Which situations did I raise when I shouldn't have? When didn't I raise when I should have? You can't get the answers from such a stupid stat. Worse, you can't get the answers even if I provided a ton more info. It is the wrong approach.

Again, this isn't me just mouthing off on a silly message board. My team has something like 20,000 hours working on this very problem. Of course, our approach works precisely because it doesn't relay on measuring your game objectively. In fact, we don't care about the user's past play and we don't provide any of the b.s. historical product features - eg., hand playback, hand downloads, comparison charts, etc. They are not needed for a player to slaughter a table.

Mathematics will absolutely not get you there. We tried in our original effort and it failed miserably.

The question of winning or losing is never asked as it is an unimportant question. The program does not track winning/losing. The program simply cares about one thing - what decision do I make now? It is all about specificity. Certain things happened before the action came to me, certain things will happen based on whatever action I take, and now I'm holding specific cards. That is all I care about. At the moment of decision I couldn't care less about my historical results as they mean nothing (other than how other players perceive me). Knowing I historically raised 20% preflop means nothings. The fact that I am a winning or losing player means nothing. The knowledge is not actionable. I have a decision to make and all I care about is making the best decision at that moment. Everything that doesn't help me make that decision is just worthless noise.

You just can't measure things objectively and even if you could it is usually of little value. This doesn't mean that results never matter. Rather, it is a softer fuzzier analysis.

It is easy to talk about objective measures in poker but people can't back up the argument with even an objective review of a simple problem. That is the folly of "objective measures".

[Utah]

Hi Cryus,

Okay, that is a lot of slicing and dicing of my comments so I am not sure exactly how to respond as there are a lot of ways to look at this issue [img]/images/graemlins/smile.gif[/img] Therefore, I will take an initial stab and please direct me if I am not answering your question.

Lets start from a most fundamental scenario - It is the preflop. You are on the button with A,8o and it is folded to the cutoff who raises. What is your play. You can call, raise, or fold.

Now, in this situation all you care about is if you should call, raise, or fold. You dont care about anything that doesnt help you determine which of the 3 options you should take. Everything in poker boils down to a decision and any information that isnt actionable is worthless.

Therefore, in the above scenario you dont care about the following:
1) Your historical win/loss
2) How often you raise preflop (other than how it affects other player's perception of you)
3) How often you call preflop
4) Any objective measure of your game

The reason is that these things cannot help you in your decision. Thus, looking at your overall historical results, i.e., objective measures, cannot really help you. Lets say that you call preflop more than the average winning player. So what? Do you start calling less then? If so, in what situations?

It is all quite silly to think that stats will help you.

Now, lets take the other player. You do care a lot about his historical play as it gives you an indication of what he holds and how he will play. BUT, what do you really know about the other player?

Lets say you are using one of these lameass stat products. It tells you that he raises preflop 30% of the time. Well, that may be a touch useful. But, that stat tells you little about what he is doing at the moment. Now, it would be very useful to know that the player raises 50% of the time in the cutoff when it folds to him and there is a certain player profile seated in button, sb, bb. but, how are you to know this player with this degree of percision? And, even if you did, how would analyze how he is going to play on the flop and beyond in certain situations? And even if you knew that you still have no idea what hands he will play the 50% of the time he raises. If you have 10,000 hands on a player you may at most have a couple of hands from this situation that made it to the showdown. Even then, you are stuck with the deviation in cards dealt. Finally, what if the player varies his play?

What you have is a bit of mess dont you. So, what players do is use intuition and starting hand charts and what not to guide them through the game. But, these are very broad strokes.

Math cannot solve this. If someone disagrees please tell me how.

Certainly, game theory can help you if you have the proper input (which you usually dont). BUT, you are pretty much screwed with game theory if you are talking more than a few players with more than 1 or 2 rounds left. for example, there are 3^169 UTG stategies a player can employ. Now, assume there is only 2 players. There are 3^169 player 2 strategies. Matlab and such can at max calculate a 1000x1000 sparsely populated matrix. The preflop matrix of 3^169 x 3^169, if calculated with the fastest computer in the world starting at the big bang, would be something like .000000000000000000000000000000001% complete by now.

Now, throw into the mix that there are 100,000s of players at a table and you rarely see the same players or have decent hand workups on them.

There is simply no way these little stat programs will hep.

There is a much better way but I will keep the IP protected for now. However, I will answer questions if I can without giving anything away.

[Martin456]

Can you give us a hint?

[raisins]

Hi Utah,

I followed this over from the Politics thread. For a long time I've found these questions to be some of the most vexing to answer. Math does give you an answer, but it is dependent entirely on the assumptions you make about multiple variables and most of the time they are completely made up.

I also remember an old post of Tom Weideman's on RGP noting that the ability to predict an opponents' actions on future rounds was a vastly underrated skill.

If I was working on this type of software project, here is how I would go about soving it. First I would datamine the game. Then I would fit the data into profiles. Determining what made a profile would be subjective and would be dependent on the ability and intuition of the developer. I would build collections of lines with the assumption that players who take the same line in a given situation would resemble each other in a different one. If this worked out you would know not only how often a PF raiser who was 3-bet PF and then check called the flop folded the turn versus check raising but also the likelihood of having the opportunity to check raise the river knowing the opponent intends to bet rather than take a free showdown when the overcard or fluch card rivers.

Probably the most important variable to focus on in creating the profiles would be not just the betting patterns but the flop textures that accompanied the patterns. If this or something close to it is what you did, congratulations, it must have been quite difficult.

However, I am a bit skeptical. The assumption that players who play one situation the same will also play other situations the same is necessary but I'm not convinced it is reliable. If it is, then I would think this would work best at low stakes games with opponents who at the most think about the hand you might have but they are not thinking about what you think they might have. In other words, the more of a thinking player they are the less patterned their decisions are going to be and the more attention they will be playing to you and their responses will be more spontaneous and irregular. Has your program been of benefit to you at levels where there are some tough players like 10/20 SH limit? I would also suspect that this approach would be of less value at NL where opponents will change up their lines depending on what they and their opponents have been doing in the recent past or what they suspect about their opponents mood. Have you used your program at NL? If not, do you think your subjective analysis would work there?

I've been paying some attention to your discussion of the software the past few months and I'm glad you chose to discuss your thinking about it a little bit. If nothing else, it was interesting to think about.

If you can respond to any of the areas I brought up while maintaining some dicretion about your work I am interested in hearing your thoughts.

regards,

raisins

[Andrew Prock]

I think you guys are confusing two very different things. While PokerTracker might be a useful tool, it is not the beginning or ending of "math". Math is very relevant to solving these sort of problems. On the other hand, PokerTracker numbers might not be that useful.

- Andrew

[ACPlayer]

[ QUOTE ]
I think you guys are confusing two very different things. While PokerTracker might be a useful tool, it is not the beginning or ending of "math". Math is very relevant to solving these sort of problems. On the other hand, PokerTracker numbers might not be that useful.

- Andrew

[/ QUOTE ]

On that we can agree. The decision cited by Utah still comes down to a decision on what is the action that leads to the best possible gain (or at least it should). However, clearly your personal win rate, has nothing to do with the decision. If you have an objective measure for the persons raising standards (I know players who raise the same UTG and CO) then use it, if you dont have it, develop it (or anyother measure that floats your boat).

Similarly, if his system is really better than what he was using in the past then this better be based on some actual hard data rather than just his opinion, if he wants to convince others that he is right. Otherwise it would be like most in the card room who reply "I am doing OK" when asked how they did -- meaning the wife has not left him yet.

[Cyrus]

I think we are fundamentally confused here. [img]/images/graemlins/smile.gif[/img]

There are two things: One is your conditional expectation, meaning what to do in the specific situation in front of you. In other words, you are seeking to maximize the chance P of your winning in A when givent the conditions B, aka Pw(A|B). The other thing is the historical overview of your performance.

The two are not the same thing - and I would categorically agree that the second kind of analysis would help you little in the first endeavor.

[ QUOTE ]
Lets start from a most fundamental scenario - It is the preflop. You are on the button with A8o and it is folded to the cutoff who raises. What is your play? You can call, raise, or fold.

Now, in this situation all you care about is if you should call, raise, or fold. You dont care about anything that doesnt help you determine which of the 3 options you should take. Everything in poker boils down to a decision and any information that isnt actionable is worthless.

Therefore, in the above scenario you dont care about the following:
1) Your historical win/loss
2) How often you raise preflop (other than how it affects other player's perception of you)
3) How often you call preflop
4) Any objective measure of your game.

[/ QUOTE ] I agree.

(But I can't help noticing that you just referred, in #4 above, to "the objective measure of my game". [img]/images/graemlins/smile.gif[/img] Didn't you just write that "We just can't measure things objectively" ??)

These "little stat trackers" such as PT or StatKing only inform you how you have been doing in the past, overall. To give an elementary example: If you have been playing in a wide variety of medium limit hold'em games for a significant amount of rounds, but nothing of the extreme kind, i.e. no games with mostly excellent players nor games with mostly bad players, and your little stat tracker tells you that you have been involved in something like 50% of the pots - then, brother, the next time you are faced with A8o that lit'l stat means nothing for sure, but when that game is over and you head home, you better sit down and do a serious re-evaluation of your whole game, because more probably than not you have a leak. Of the loose kind. You probable need to tighten the screws, some.

(Notice how cleverly I slipped the words "screw" and "loose" in our dialogue. [img]/images/graemlins/smile.gif[/img])

[ QUOTE ]
What players do is use intuition and starting hand charts and what not to guide them through the game. But, these are very broad strokes. Math cannot solve this. If someone disagrees please tell me how.

[/ QUOTE ] When I wrote above that the 50% participation in flops is probably excessive, that is not based on intuition or any "broad stroke". Under the conditions I formulated for your record (many games, good games, nothing excessive, limits) you most probably have negative EV by seeing half the flops. And this is not intuitive.

Now, I have no idea how to put together a poker bot but in the scenario you gave (A8o on the button, folded to CO who raises), there are very specific parameters that can be entered and be helpful in the assessment of the situation, i.e. searching for the what-to-do. Such as, the perception of other players abt your playing style (which, please note, takes into account your stats in that game, tracked via a ...little stat tracker); the other players' style (ditto), if possible over a longer timeframe than just that particular game; the mathematical expectation of an A8o aganist random hands to the river (derived from sims); the size of the pot already; etc etc.

But I can't help thinking that we started this whole thing on a big misunderstanding.

--Cyrus

[Utah]

I dont think there is a misunderstanding......

[ QUOTE ]
I think we are fundamentally confused here.

There are two things: One is your conditional expectation, meaning what to do in the specific situation in front of you. In other words, you are seeking to maximize the chance P of your winning in A when givent the conditions B, aka Pw(A|B). The other thing is the historical overview of your performance.

The two are not the same thing - and I would categorically agree that the second kind of analysis would help you little in the first endeavor.

[/ QUOTE ]

All you EVER care about as a player is Pw(A/B). It leads to specific decisions. So, you are exactly making my point - historical overview cannot help you [img]/images/graemlins/smile.gif[/img] So, why ever do it?

[ QUOTE ]
there are very specific parameters that can be entered and be helpful in the assessment of the situation

[/ QUOTE ]Please give an example. Those factors are wickedly hard to figure out in the online world. And, even if you could figure them out, there is no formula you can plug them into mathematically to derive the solution, expect for maybe a 2 or 3 player game on the river.

[Utah]

[ QUOTE ]
Math is very relevant to solving these sort of problems.

[/ QUOTE ]Yes. What I mean is that the whole problem can not be solved directly by math. Certainly, things like calculating pot odds, implied odds, % chance of being called, etc. are very important.

[Utah]

[ QUOTE ]
On that we can agree. The decision cited by Utah still comes down to a decision on what is the action that leads to the best possible gain (or at least it should). However, clearly your personal win rate, has nothing to do with the decision.

[/ QUOTE ]Exactly my point. If it doesnt help you then it is just noise. As a player I only care about maximizing each Pw(A/B). Thus, I never ever need win/loss stats.

[ QUOTE ]
Similarly, if his system is really better than what he was using in the past then this better be based on some actual hard data rather than just his opinion, if he wants to convince others that he is right.

[/ QUOTE ]I already convinced others I was right and raised a lot of investor funds awhile back to build it [img]/images/graemlins/smile.gif[/img]

The original discussion on the other thread wasnt meant as a, "I have invented something better" discussion. I also didnt say data couldnt be used. Rather, I said it couldnt be solved mathematically. The info is fuzzy so we use lots and lots of neural nets instead. Now, that is nothing totally new in itself. However, we use them in a unique way and we dont suffer from some of the limitations that other net approaches suffer from.