Poker-AI.org

Although I am a noob here, but I have been botting for 10 years. I can document that later.

I use a tool that allows me to simulate specific NL Holdem scenarios, a statistically significant number of times, in order to determine MaxEV for each. I then record and store the action to be used when it arises in tournament play. While the number of scenarios approaches infinity, I assert that there is a way to break the game up into distinct scenarios such that there is a real, albeit huge, number trials to be tested.

It is a fact that there is a perfect way to play blackjack sans card counting, etc. For instance, although it is a losing proposition, always hit 16 vs. a 7 showing. The casual player may argue, buy if you do this a statistically significant number of times you will determine that this will MaxEV.

I contend that NL Holdem can be divided into blackjack like scenarios. A very simple example: hero is in the 2 seat QTd. The flop comes Qh Th 9h. There are 5 players left in the hand. Your stack-to-blind ratio is 55.16, stack-to-pot ratio is 0.44, and your stack-to-opponent ratio is 2.03. Among the players at the table, you have the biggest stack. The first seat limps. The pot odds are now 12:1. What action will yield MaxEV for hero? Clearly there are more variables to be considered, many are taken into account but not listed in this example in the interest of brevity.

To determine the best action, I create a ceteris paribus, all things being equal, simulation and vary only the actions of hero. If you do this a statistically significant number of times, you would determine that in all cases, you would MaxEV by calling. No surprise there. Now lets complicate things. Same situation only, but now hero is on the button. There are now a huge combination of scenarios that can confront hero. For instance, what is the proper course of action if it is checked to him, player 1 limps and all call, or player 1 makes a pot sized bet and there are two callers, or player 1 bets and is reraised all-in by another player without a caller before hero, or myriad but finite other possibilities.

Here is where I differ. Rather than create an algorithm to determine the best action, I break this scenario into a discrete number of possibilities that are predefined. In the example rather than evaluate the bet in increments of 1 chip, I define the possibilities that hero may face as No Action, Small Raise, Normal Raise, Raised Back, Big Raise, or Huge Raise. Each of these possibilities can be expressed as fractions or multiples of the pot, number of bb, or fractions of hero's stack. I then break down the hero's possible actions vs. any of the above as, in the instance of a Small Raise, hero might Call 2 or fewer players else fold, or Fold 3 or more callers else, or Raise exactly 1, etc. In this case there are 18 possibilities. The size of the raise is determined in increments of either pot, stack, or bb.

Here's my problem. The tool I use is not mine, and it is no longer available. I have spent literally thousands of hours over the past 10 years manually making each change, simulating it, then changing it again until I can achieve MaxEV for each. However, I believe that I have only scratched the surface of what is possible. I need help automating this process.

I am very interested in your opinions and questions. Thank you for reading.

Hi Jester!

On what limits are you botting? 'Cause as I understand, you don't use opponent modelling?

About problem, if your approach is working good for you I think you should create that tool for yourself, so your developments can continue to work.

"manually making each change" why manually? There are many simple algorithms for tuning parameters like genetic algorithm.

You sound as if you know what you want. What's stopping you making it?

If you have been using the same tool for 10 years I am gonna guess it something like Holdem Inspector as that was really all that was available back then. If so I have a DLL you can use that will give you identical results. I even have the source code somewhere. I didn't write it but a friend of mine did when I was running limit bots 10 years ago. Also..why cant you just use the old version?



Well..this may spark some controversy but I really do no think there are that many possibilities just in the fact that your going to need a hand in order to beat multiple players. Top pair would even be a marginal hand with a bet and a call, two callers etc. Bluffing multiple players is obviously not going to work very often either. So most of the time there is never much of a decision to make unless you have a big draw or two pair or better. HU is where it gets complicated in my opinion. This is why an aggressive preflop isolation strategy is generally the best one to adopt so you can avoid multiway pots. However..with that being said using a small ball "calling" strategy can also be effective. Both have their merits but in the end I think the aggressor will take down the most money long term.

your algorithm that you've used for 10 years doesn't work because as you do more iterations, the opponents with weaker cards than you would learn to fold immediately since you are simulating it with your cards being identical each time.

Thanks for the feedback.

@ Dowakin - I started grinding limit games .10/.25 using 5 servers each running 4 tables and switching shifts every 8 hours. I was grinding out about $50 per night on average. It sounds like easy money, but it was a full time job having to chase down hiccups, reboot, switch tables, etc. So started selling my profiles, and sold nearly 1000 profiles over the years.

Regarding opponent modelling, I had the ability to track all play into a central PTrack db, summarize the stats of each opponent by the basics, assign a category (Rock, Solid, Typical, Fish, etc.) and act accordingly. Honestly it may have been a limitation of oversimplifying the analysis, but I never found a difference in real world results and turned it off.

Overtime I proved to myself that if something works better in the simulator, it works better in practice. At first I would use online sites as proving ground for the profiles, running 10k hands and measuring the results. I know that is not statistically enough, but I gave me a good feel for the profiles. Once I was certain of that, my emphasis changed from botting to creating simulation based profiles. I haven't run a bot farm, or played limit poker in years, but continually improving the simulated results of my profiles has become sort of a hobby.

Then came the NL program. I started all over again. Not botting per se, just playing some automated speed poker, and sngs. I never sold any of the NL profiles. Now I use them to help me grind out bigger tourneys.

My problem is that the software is proprietary and defunct, I can't get the former owner to contact me even though I was the top contributor on his site. Even if it was possible to modify it (is it?), I have never written a stitch of code in my life. I am just an economist that plays poker a lot, pretty well actually. That's why I decided to show up here.

I am not sure how you use "algorithm," but I think that is where we differ. For those unfamiliar, a Monte Carlo measures the results of a statistically significant number of trials and determines the probability of different outcomes. To me, an "algorithm" calculates the probabilities using statics and math. For example, it is simple to determine that the odds or rolling a twelve using two standard dice, 1:36. Using a Monte Carlo, you would roll the dice a significant number of times and count the results. The result will eventually approach 1:36.

Applied to poker, I try to determine the maxEV for a hand in each and every situation and always choose the play will maximize my chances. A simple example is AA from the button when it is folded to hero. Hero calls, always. That's where it gets controversial. Mock me if you like, but I can quantify it. The trick breaking down the game into scenarios or "buckets" (credit to Nasher) for each and every possibility, such that the sum of the buckets represents every possible scenario. It's possible and has been done.

Now for the reveal. I am using Online Poker Inspector or OPI. I waited to tell you since like Rodney Dangerfield, it gets no respect. Probably due to the juvenile stoplight in the corner which is supposed to signal what to do. I am seen it reviewed as average at best, a training tool at worst. None ever mentions the simulator. If you look it up, you'll find that the website is down, but you can still find demos on CNet. The sim is disabled and it is, due to the lack of updates, becoming less and less compatible with most sites.

@shalanko - Respectfully, we could argue the merits of any particular strategy, but we will never really know who is right. I argue that the criteria for being "right" is the maxEV as determined by simulation. I know that if you, simulate 10k tourneys with 800 players in OPI, keeping everything else the same, including the cards and the opponent play, you will achieve the maxEV by calling there. I completely agree that the micro level there may be things we humans and maybe ai can do to give you an additional edge. My point is that eventually, all that averages out. What is left is the best way to play each player, so as to maximize the potentiality outcome.

@algonoob - It works. Where's my motivation to make this stuff up?

This should shake things up a bit, but I will leave you with this. I think that opponent modeling is irrelevant in the aggregate. If you do something enough times, facing enough opponents, you are left with the best way to play on average, which is by definition, the best way to play long term. Firestorm to follow.

I have some old video and stats on my site. http://www.jesterpoker.com

DO NOT BUY ANYTHING FROM THERE.

The program associated with the profiles is no longer available which renders the profiles useless.

@Jester - I'm not claiming you're making it up, just that from a game theory perspective, the strategy you're calculating is meaningless. Any positive results you achieved is merely a result of very poor play from your opponents; I see on your site you only have stats for 0.5/1 and 1/2, and even at the low stakes of 1/2 you are already only breaking even.

You're looking at the limit stats from years ago. How is the strategy that I am calculating meaningless from a game theory perspective?

Hey..look at my post above..that is what I thought you would be using. I have a DLL you can have that will give you identical results to OPI so you can continue with your work. I still use it for "absolute equity" but have since migrated to range equity entirely. Anyway..if you want it just PM me and I will dig it up for you. Also..I believe the Timmys hand evaluator available here will give you the same results along with hand state stats with minor tweaking. There is no reason to keep using the original OPI.

I remember that stop light from years ago all too well. My first limit bot on PP used it. Automating PP and OPI at the same time became slow and cumbersome so a friend developed a DLL for me to give me the same results. You just pass your hand, board cards etc to it and it will return the equity.

When running my limit bots from 2002-2006 I played primarily .50/1 thru 3/6. I was winning 5BB/100 at .50/1. At each level above .50 the win rate was cut in half. In order to compensate for that I added in player modeling which slightly increased the win rate but not by much. I found all the work I did for modeling did not prove to be worth it for those limit games at the time. It definitely needed more work and I had better things to do. So I gave up at 3/6 and stayed in .50/1 and just played more tables modifying the AI to play tighter games which my bot was avoiding. I was more comfortable with the swings at those limits anyway. I think on a normal day I was playing about 4 tables during the day and 10-16 at night when the action picked up.

Wow, 5bb/100 is excellent. Maybe your method yields itself opponent modeling. Maybe mine would too, if I gave it enough of a chance, or maybe if I created better modeling criteria. My opinion is more quantitative than quantitative. I think that we are agreeing that if there is an edge, it is not a huge advantage. That being the case and with the high standard deviation that we all experience, I think that it would take a very large sample size with both modeling turned on and off, assuming no other changes were made to the bot, time of day, day of the week, etc, to find a difference. That could be said for any change you make.

That is another reason I rely on the simulator, and if I am right, simulation may be the only way to analyze minute tweaks to any profile. I never had the patience to make a change, play 10k hands, evaluate, and do it again. Even if I did, I now don't believe that 10k hands are not sufficient. When I run a sim, the results for only 10k hands, changing only the cards, will yield results that can vary by plus or minus .5 bb/100. Still way too much variation to evaluate even large changes.

For the record, being a laborious process. I don't think that I ever scratched even the surface of the potential for using big Monte Carlos to determine EV. I spent most my time evaluating different Pre-Flop scenarios, and changing "global" settings that effect the greatest number of hands. Is there a link to Timmy's work that you referenced.

I know I tend to ramble, but I am enthusiastic about the group getting a better understanding of my process and vetting its opinions on its merit rather than my ability to explain it. If you, or anyone else is interested, I would enjoy setting up a WebEx to show what I am doing in more detail. I doubt that I am doing it justice here, and I think you would be impressed.

My goal is to get a convert who can dedicate time to develop a way to automatically change the variables, and to save the best results. I have no intention of selling anything anymore, and honestly I am really uninterested in botting online except for testing. My passion here is academic more than profit driven.