Experts Vs Computers Vs Documentators

I want to clarify my stance on this issue. Specifically when it comes to rating the EV of a particular hand in certain positions in various type games. The question is, who will get closer, an expert player or a machine.


The reason why this is a confusing question is because there are two types of experts and two types of machines. If Roy Cooke is asked how much JTs on the button is worth on average, I would expect his answer would mainly come from his estimate of how that hand has done for him over the years. And there would be three sources of innacuracy there (Four if the sample was small). One would be his memory. His estimate of his own results may be wrong. Two, is the fact that other players might achieve different results. Three, is the fact that he is only thinking about his own game. Notice that the first innacuracy would be eliminated if he had meticulously kept track of his results. So if he was playing online all that time, and a machine had indeed been tracking him , that machine would indeed be more accurate than Roy himself about how much he averages making with JTs on the button.


The same of course would apply to me. Even when I am only talking about my own results with AK utg, I could be off a bit merely because of faulty memory or psychological bias. And a machine that had documented my results would not be off at all. However, unlike most other expert players, I also do calculations to verify my seat of the pants estimations and will alter such estimations if the math indicates I should. These new estimations will still not perfectly match the documentation machine, but they should be close. And in fact, it is even possible that math aided estimations will be better than a documentation machine's results when the sample size is small. See why?


Now what about if I am estimating the EV of a hand for others, in games unlike the ones I play in. Here I have no chance to do better than a documentation machine with oodles of data, regardless of how good my judgement or calculations are. If I say that I think that two aces has an EV of 4 big blinds under the gun in a 10-20 game on Paradise, bet against me if Angelina says it is 5. A computer has documented the results of millions of hands. End of story. The debate comes in when the computer is not documenting real hands but instead documenting hands that it is also simulating. It's perfect accuracy in documenting results is irrelevant if it is not simulating real world play properly. This source of innaccuracy may be greater than the innacuracy of an experts judgement or memory, especially if that expert is using math to minimize mistakes.

Of course you're right.


I agree with every word you just wrote.

What you write is true, as far as it goes.


The simple truth is: it depends.


Game conditions tend to dominate when it comes to coming up with proper estimates of "EV" for situations like AKo under the gun. 15/30 Paradise is nothing like 20/40 Paradise, which is nothing like 15/30 Bellagio, which is nothing like 30/60 Bellagio, which is not the same as 20/40 Lucky Chances or the 15/30 Oaks club, which in turn differs greatly from 15/30 Madison WI.


And all of those differ from 10/20 TTH.


The real questions is, what are the qualitative and quantitative differences between the game conditions, and how can we make best use this information to make qualitative and quantitative comparisions between two different games.


[I realize that I've oversimplified things a bit, the game conditions of the 15/30 Bellagio are not static. Like all games they vary with time of day, day of week, and week of year.]


It might be possible to argue that the game conditions of TTH are so far removed from the game conditions of live play that they make any results meaningless. To date I have seen no one make that argument quantitativly. I've seen David and Mason make that argument qualitativly. But it's going to take real numbers to convince me that tools like TTH are not useful.


When I see reports that historic data from Paradise is coming up with numbers similar to what TTH produces, it tends to make me think that TTH is more useful than not.


- Andrew

Analysis of hand records, TTH sims, mathematical analysis, memory or poll 2+2/RPG/(experts) is wrong, use and.


One does not investigate a problem by deciding on one method of investigation and by ignoring all others. Instead you analysis the problem from all angles, if they all give the same result then you are likely on the right track. If there are significant differences between results then the problem should be examined further in order to understand where these differences are coming from.


Using TTH sims correctly can never hurt and can only gain. David knows this of course, so maybe he is aiming his comments at people who will not be able to use them correctly. Certainly using a tool like TTH badly at the expense of other forms of analysis could be quite damaging.

The point of all this discussion is to learn how to choose the best play, right?


From this perspective, David is perhaps overly generous about the "accuracy" of online record-keeping. Just because comprehensive records show that, say, reraising an UTG raise from the small blind with a given hand has delivered statistically significant higher profit than calling with that hand in the 10-20 game on Paradise, this does not mean that reraising was the best choice each and every time. The best choice still depends on the raising requirements of UTG, how well the big blind plays, the style of the small blind from the flop on, how much experience UTG has vs. the small blind, etc., etc.


And even if we can assemble a sufficiently large sample of hands for the precise situation (down to specific players), and reraising still shows a statistically significant higher profit compared to calling, this still does not demonstrate that *the next time* this precise situation comes up, reraising will be the best play.


The ultimate reason why so many debates about how to play a given hand in a specific poker situation cannot be resolved is because the expected profit of each choice (bet or check; raise or call) is fundamentally uncertain. This is because we cannot know for certain how our opponents will play in the future (regardless, for example, of extensive past experience).


It is possible to precisely compute an upper and lower bound for each choice. However, it is almost always the case that these bounds allow us to choose either action over the alternative. For example, the upper bound for reraising an UTG raise from the small blind with a given hand is likely to be higher than the lower bound for calling, and vice versa. (One occasional exception is being last to act on the river after everyone else has checked, when holding the nuts.)


This fundamental uncertainty is not just an "academic" issue. All of the (reasonable) debates about how to play a hand that we see on this board boil down to disagreement about more or less plausible assumptions about how the opponents can or should be expected to play.


There is a way to tackle this impasse, but it does take some effort. A minimal requirement is consideration of how one would play *all* hands one might hold in a given situation, rather than how one would play a specific hand (the latter being nearly equivalent to debate about the number of angels that can dance on the point of a needle).


For example, instead of discussing how one should play 77 UTG in a No Limit Hold'em Tournament, the discussion should be expanded to a debate about strategies for playing all 169 different hold'em starting hands UTG. The uncertainty about expected profit for such strategies is much narrower than the uncertainty of expected profit for individual hands. Thus, demonstrating that one strategy is superior to another often becomes feasible without complete consensus about opponents' strategies, and can sometimes even be achieved despite total disagreement (about opponents' strategies). In the process, one can gain insight into how your opponents may be thinking, and how you can more accurately identify your opponents mistakes.

This is the best description I've ever seen yet of the computer's powers and limitations, in simulating the game of Poker. With some expansion and elaboration, I'm sure it will find its way in a Sklansky book.


(One not yet published I mean! I am a Twilight Zone fan but within reason.)

It is well written and accurate, but I would hesitate to call something that obvious, brilliant.

What seems obvious to you (and me) is most definitely not to most others. Hence, the endless (and pointless) bickering between computer freaks and leathernecks.