Quote:
Originally Posted by MarkD
Is there a paper you can quote that your algorithm uses as inspiration or is based on?
We use CFR as our main algorithm. As you probably know, there are multiple implementations with small differences. We haven't innovated here, we tried several variants of CFR (Montecarlo, weighted, Pure, CFR Plus, each of them with their own parametrization). However most of those variations of CFR were developed looking for faster convergence, but in our case we were interested in tree completion, to minimize the number of nodes without a solution, even if there was a mistake in a previous street.
Regarding abstractions, we use an in-house developed innovative system with two sets of buckets. In our benchmarks it proved to be a least 3 bb/100 better than the buckets created with M Johansson's (2013) method (KMeans / Earth mover's distance).
The other innovation that our solver has, is to use multiple sets of buckets in different parts of the tree. Instead of using a single set of buckets for the whole tree, we use finer-grained buckets (through Elbow and Sihouette statistical methods) in some places, allowing us to give better solutions on the most common spots.
To test the quality of the strategies we use exploitability metric based on LBR (Viliam Lisy 2016) which is a good metric to get the sense on when the tree has converged, but it has limitations when it comes to compare different abstractions, that's why we developed the testing framework.
However, please feel free to test our solutions and try to find any serious flaw on them. We will give away 3 months of our Elite plan to any user reporting one. Our cost and time to re-train the solutions is much lower than traditional solvers, so we will fix them easily.