Question about TTH

I have only played an earlier version of TTH. I hear the new version is better. However I am curious if it makes a critical error. One of the hallmarks of better players is that they rarely overcall with a mediocre hand even though they would call with that same hand if it was heads-up. This is especially true on the last two betting rounds. For example, if the five cards on board were JT864, QJ should almost always fold to a bet and a call.


The reason I'm asking this question is because we are being told that TTH rates big pairs higher than we do (ev wise). The simplest explanation would be because TTH players play poorly which would tend to mean they lose too much to a pair of aces. The most obvious mistake would be too much overcalling which would tend to inflate the aces profits when they bet for value (and get two or three calls when they should only get one). Thus my question is does TTH seem to properly take into account an earlier call? In other words does TTH instruct its players to fold fairly good hands in this spot as it should?

This is a good question and I don't have an answer for you on this situation but I will say this: I have it set to be about as tough a game as possible and all the "players" call WAY too much with gutdraws. I have seen calls on the turn getting like 4-1 to try to hit a gut draw- sometimes without even having an overcard to go along with it. I don't pay that much attention when I play because I am with you when you say that computer simulators aren't worth very much. I just play to pass the time when I'm bored. Computer games are just way too different then real life play.

BTW I have the 3.0 version of TTH

I've read the most recent go around about the sims with interest. In previous ones I've stated that benchmarks would be very useful in evaluating the effectiveness, reliability, and correctness of the software. IMO this could be used as one benchmark and I think it's a good example of one. Sorry I can't answer the question.

I am not sure if this will help, but it may give some insight into the profiles.


If you set up a simulation giving the calling station (simple simon) profile AA - and pitting him against 9 tight aggressives - you will find that his check call method of play is far more profitable than the play of a tight aggresive with AA against 9 other tight aggressives.


I think that this may be because the tight aggressives over rate top pair post flop and, by playing AA passively pre-flop, the calling station profile allows them to stay in the hand and hit top pair.

Mason,


You ask a really insightful question.


And I learned that this was an important question only after writing my poker playing program modelled after TTHE and playing it for many thousands of hands at U of Alberta. Before I answer, I want to say again that Wilson's playing encodings are really really amazing, they saved me a huge amount of work.


But your question, does TTHE do it? I do not know for sure but I can make an educated guess 'no'. I know that I cannot use his setup screens to set this up. Simulations I have run look like 'no'...I therefore assume it does not. But Wilson can tell you for sure. Wait, stop the presses, in rereading my answer, I see something: TTHE does pot-odds. He must know how many are in now, and how many might stay to do a proper calculation. Maybe he does take it into account for other things... hm....


What a program should do:


My program's win rate at the Alberta poker server, which is not an easy place to win, was going nowhere (early on) until I ran enough hands to realize a number of things you cannot encode in 'tables':


1. You must always take into account how many are left. In situations like this one you describe, in bluffs, in pot-odds (where you can calculate how many are left, so what you might win if you hit), you must know. But interestingly, especially in bluffs/semi-bluffs (encoded as X1 by Wilson), and calling with second pair, etc, you better better better better know.


2. You better bluff. (I mean a program better bluff). Wilson does this through encoding some bluffs, but it is not enough. My program bluffs randomly in certain other situations...nuff said, don't want to give too much away.


3. You must always play some situations very aggressively, tables be damned, and play other situations very very meekly. Someone above mentioned gutshots. I found a pattern -- which I know TTHE does not adhere to in its tables for its best players -- independent of tables, which describes how best to play low end str8s and gutshots -- which is almost never. Perhaps I misread the table encoding, so Wilson, if you are reading this, no disrespect is intended.


There is much more, but it is tough for Wilson, who is selling off-the-shelf software, something I know all too well about, to program certain things into his product in a very general way, for people to encode.


I would not want to play against Wilson's REAL program, believe me Mason, he knows how to write the real thing.


And I would not like to play against mine either, and work is not even half done. Yeah, it plays real well, but it is just a start.


Mark

For example, if the five cards on board were JT864, QJ should almost always fold to a bet and a call.


To answer this question specifically, all of the profiles that I tested MADE this call in the particular scenerio that I simulated.


That's just tip of the iceberg. There are worse errors than that.


Keep in mind, however, that the TTHE would become far more vulnerable to manipulation if they were to fold too often, than the other way around. It's probably better to have them call too much on the river than fold too much.


In order to properly use TTHE, IMO, you have to modify the simulations to help account for some of the errors that they make. Keep in mind, that these errors go both ways, and sometimes AA puts in too much money. So its not all a one way street, although its likely that AA is benefiting more.


If you will notice Sklansky's UTG EV estimate was $1.25. Mine, using modified TTHE sims, was 1.20 (including rake/toke). My AA UTG was making $90. Sklansky's got it at 50 and seems to be willing to go higher, based on his answer. So my modified TTHE sims are showing the pairs (and AKo) still higher than Sklansky's baselines.


Do the modified/controlled sims better approximate human play? They seem to, but still not fully, in my view. I don't have the time and inclination to properly document for others what the errors are, and how to modify for them...so just take this as speculation.


Are these modified sims better than the more generic ones? Not necessarily better, they just have different underlying assumptions about how the opponents behave.


Keep in mind that the users of TTHE generally know its limitations better than its critics.


Regards.


ps I wonder if these differences in these EV estimates account for the slightly less agressive play with big pairs by S&M that others seem to advocate? Or is that neither here nor there?

I don't think there's a "real" version of the program that plays much better.


Here's why.


TTH version 4 plays quite a bit better that vers 2 or 3

and one new feature that's sort of fun is the

"Challenge Mike" contest.


First the resident expert "Mike" plays 50 or so hands and keeps

his plays and results a secret.


Then you get to play the same 50 hands and see how you did.


Now, if there were a real version that could play much better,

there would be no real reason not to use it as Mike's

hidden strategy.

Hi Mark!


I have been following your project off and on for a year-and-a-half or so, and I have thoroughly scoured all your papers and waded through much of the code.


How much time do you devote to that project and how's it coming along? I haven't seen a new paper in a while, so I reread the ones for inspiration. Actually, I would like to develop a similar program just to say to myself, "I did it!"


The last published paper, I believe, dealt with improved opponent modeling and used neural networks as a non-real-time method of identifying opponent playing styles. It was mentioned that current CPU speeds were not able to cope with doing neural network calculations in real time.


I have read a couple books on neural networks, but have not implented them in programs of any complexity. What CPU speeds would be required to run these NNs in real time?


ThanX

John,


You got the wrong guy, you want Aaron Davdison of U of Alberta.


My project is only 2 months old, and I use a different method than Aaron and his team...


Mark

Why not just pit the TTH profiles play against cash games online and see how they do over a couple of thousands hours at different limits?

AAAAAAAAAAAA!


Cool...do you have a website following the development of your program? I'd be interested in reading about it!