How is Deepsolver so good?

How is Deepsolver so good



Back to help center

Deepsolver Process Explained



First commercial Poker Solver powered by Neural Networks - this is how it's done.

To understand how our technology will translate to your edge at the poker tables, first, we need to take a look at how old generation solvers dealt with the task.

Get up to speed



Old solvers are usually based on CFR algorithms. How do they work?

As a standard, CFR algorithm has to create a whole decision tree for the game. The algorithm then runs (iterates) every possible scenario for every consecutive street: flop, turn and river. Old technology has its limitations though, like expensive hardware, time-consuming setup and low pace of the calculations, which might be prohibitive for most of the players. It's XXI century, after all, no need to wait hours for your bulk calculation results.

As you can see, the old way is not the most effective and flexible solution out there, resulting in slow operations and prolonged waiting times for proper analysis.

Our algorithm is more of a hybrid, combining our proprietary CFR Engine (for calculating current streets) and Neutral Networks (for estimating Expected Values of each action at future streets). That allows us to build much smaller decision trees than our competition, resulting in dramatic speed improvements (check our Benchmarks and Testing blog post here).

The New Way. The Hybrid Way.



To answer the need for poker professionals, we decided to turn the tables and think of a new way of doing things. This is where our Neural Nets come to play. Our approach is different. First, we took a deep dive into AI & ML research papers, creating the first commercial Neural Net-powered poker solver in the world. This is how we do it:

Deepsolver combines the accuracy and precision of CFR algorithms with the speed and agility of our Neural Networks. Thanks to that we are able to solve one street at a time, instead of building whole multi-street decision trees. This is possible due to our Neural Networks being able to estimate EV at the terminal nodes of the tree. We don't have to calculate whole decision trees every time with every variable. We don't lose time to check nodes that make no sense in real-life scenarios. For you, that means unparalleled speed and super-human precision.

Example:
If you want to calculate flop we use the CFR engine to calculate flop, but we use Neural Networks to estimate EV of the turn and river at their terminal nodes. Thanks to that we DON'T HAVE TO build whole decision trees with every possible scenario.

Neural Networks Training



To be 100% sure our Neural Networks are able to generate proper EV at the terminal nodes of the action, we invest a lot of resources in their training process.

Our Neural Networks were trained with 4 sizes (including small bet and overbet) to make sure EV estimations on the next streets are as close to real-life EVs and scenarios as possible.

Until now we generated millions of turn and river examples to train our Neural Networks. We used real-life scenarios, various sizings and a wide library of ranges, both real-life and corner-case examples.

To train (or teach) our Neural Networks we deliver data in packages and give it a task. Then the Network is transforming to match data with the results, perfecting itself every time it does that. As we used millions of examples and data points, we're confident that our Network can deliver both the speed and precision needed by poker PROs.

Limitations of traditional CFR:

not effective - runs through every place, every node and every branch of decision tree
sloooow - as it has to run all over the place, the speed is just not there
hardware required - you still need to buy private servers in XXI century
not flexible at all - no matter texture of the board, next streets will be calculated with the same sizings (you can't change them) and with limited decision tree depth

Limitations of Deepsolver:

not able to calculate exact Nash Distance of given strategy in real-time
not able to conduct multi-street toy games experiments due to fixed 4 sizings used as betting options for each player at the next street.
Neural Networks have a certain loss, which may result in not being able to converge to super-low values

Back to help center

Updated on: 21/09/2021

Was this article helpful?

Share your feedback

Cancel

Thank you!