Poker-AI.org

I am at the point in my HU bots development that its nearly ready for live testing. I was concerned about many aspects of the bots long term actions as many things are just not obvious until I get enough HH files to analyze. Since that is a slow process I decided to write a program to test the AI for leaks in various situations. Two areas I needed to concentrate on was how often it cbet the flop and how often it folded the turn vs a barrel. For the baseline strategy to be effective and for the most part unexploitable I had to make sure these two aspects are leak free.

So I wrote a program to test each of those situations. The first situation is the bot raises preflop, villain calls and the bot leads the flop. After running thru 25k boards the stats stabilized and I found that it was not cbetting enough and that is was value heavy. The toughest villain cbets around 45-50% so I adjusted the bluff percentage to nearly 20% to achieve balance. Roughly it turned out to be 18% value, 18% bluff and 15% checkback/call and of course 50% fold. Of the checkback range 50% of that was value to remain polarized and to keep the villain from always betting the turn and making my hand transparent. I think this is nearly correct but I may need to widen the checkback range. Of course this all goes out the window if the villain folds to more then 55% of cbets as I have it widening its bluff range to take advantage of that tendency.

So now I use the LeakFinder to test all aspects of my AI to ensure its playing correctly. This has been an incredibly useful and valuable tool and I highly recommend writing one to specifically target certain situations and not just the strategy as a whole.