*First I want to preface this by saying; Even if you don't play poker, there is some interesting stuff in this post that can apply to other strategy games. So if you enjoy strategy games or you have an appreciation for game theory in general, I think you will enjoy this one.*
---
Today I would like to share a discovery I made in the last year that opened my eyes to a level of complexity I didn't even think was possible with the game of poker. Join me today as I dip my toe in the world of solvers and AI in poker.
In this post I look at what I always thought was a very straight forward spot in No-Limit Texas Hold'Em. By the end of it, I can say that it completely changed the way I look at the game.
### A Bit of Poker History
Before I get into the meat of the subject here, let's talk about the history of poker strategy a little bit...
Back when I started playing, the best way to learn the game and create winning strategies was to:
- Read Poker Books
- Use Mathematics and Game Theory
- Compare strategies with known good players
Obviously these are still good ways to get a better understanding of the game if you're just starting out. But it was all based on theory and reviewing past hand histories. So there was always some level of uncertainty when you were putting together a strategy. In other words, it was pretty much impossible to know if what you were doing was optimal at all.
Sure, you could use probabilities and simple math to calculate *Expected Value* for a given action, but you couldn't definitively predict if your newly crafted strategy was going to work in the real world. All you could really do is try it for a while, and analyze your results to try and adjust where you could.
### Poker Meets AI and "Solvers"
Now, we have powerful tools that we can use to run simulations and test out our strategies in a matter of seconds (or minutes, depending on the computing power). Solvers can evaluate all possible actions in a given spot and tell you exactly what the optimal strategy would be, they can achieve that by testing all possible scenarios (yes, all) within a set of constraints that you give it (decision tree¹). This allows us to discover errors in our strategy that would have been very difficult to spot, or would require a very large sample size to notice without a solver.
### The Strategy
This brings me to a major leak I recently discovered in my strategy for No-Limit Hold'Em tournaments. More specifically, I'm talking about my strategy for playing out of the *Small Blind* when everyone folds and it becomes a *Small Blind* vs. *Big Blind* situation.
See, traditionally it was believed that you should mostly be raising or folding and try to avoid limping or calling from the SB. Sure there are situations where flat calling could be okay, but generally speaking, it's always mostly been fold or raise.
But I recently noticed that my win rate in the *Small Blind* was lower than it should be, so I was looking for ways to improve on that and here is what I discovered...
The solver **NEVER** folds in the *SB*, it either *RAISES* or *LIMPS* (flat call). Sounds crazy? Well it's not!
Let's take a look at my older strategy and compare it to what the solver would do.
**(Red = RAISE, Pink = FOLD or RAISE, Green = CALL, Blue = FOLD)**
|"Human" Strategy|Solver Strategy|
|-|-|
|||
### The Obvious
The most obvious difference here is the solver tells us we should NEVER open fold from the SB when folded to us in a tournament (or a cash game with antes). The solver will play every combo either as an open raise or a limp.
It becomes really obvious that we were playing way too tight when you see the range the solver wants you to play. It was a much easier strategy to memorize, of course. But it is far from optimal.
### Balance is Everything
The less obvious but most interesting part is how the solver balances for the fact that it's playing 100% of hands by using a polarized range for raising as opposed to the more linear ranges that have been commonly used by humans for the longest time. It's not as intuitive or easy to memorize of course, but what it does is it's virtually impossible for your opponent to exploit you when you're using a perfectly balanced range.
The solver picks a few strong combos and uses those as a call. It also picks some weaker, but very playable combos as a raise. The result is that whenever you raise or call, you're not giving any information as in either cases you're gonna have strong combos and weaker ones as part of your range.
### Conclusion
With the advent of modern AI and solvers, it becomes clear that some of the things we've been doing for years that we thought were correct, are actually very suboptimal. Solvers are able to create much more complex strategies than any human could with pure math or game theory. Modern computers are able to run through every single possible scenarios at a decent speed and give you a definitive answer for every spot. This sometimes creates situations where the solver strategy is so complex that it would be almost impossible for any human to execute it perfectly. So the most important skill to have now is to find ways to simplify those strategies and apply them in a way that makes sense for a human player.
This was just a simple example of one of the most straight forward preflop spot in Hold'Em, so you can imagine how complex it can get for more unusual scenarios with more than 2 players involved.
Nowadays, the best players all study using solvers, and the ones who are able to best apply those newly found strategies are the ones who will stay on top of the game for the coming years.
---
*Thank you for reading! I would love to know what you think about AI in strategy games and what the future holds for games like online poker. Please share your thoughts in the comments. 😃*
---
My last post: [About time! My plans for content in 2021](https://peakd.com/hive-169926/@chiren/about-time-my-plans-for-content-in-2021)
My VIMM Channel: [Where I stream most of my poker sessions](https://www.vimm.tv/chiren?ref=chiren)
hiveblocks