Quote:
Originally Posted by ZenFish
How do multiple post flop bet sizes affect convergence for pre flop solving?
When solving pre flop with PioSOLVER, multiple sizes are not recommended since pre flop EVs won't change much, and the solver can struggle to converge when multiple size are used (especially the ALLIN).
Does the same apply for MonkerSolver? With default settings the AUTO tree building options insert multiple sizes in lots of places.
Quote:
Originally Posted by samooth
these questions feel a bit weird to me; you are using a solver that uses bucketing (along with some suit "concessions" for plo) to reduce the game size. while it is known that bucketing is lossy, it's tough to say how lossy it really is. the only (public) info on monkersolver is a (preflop chart) comparison with cepheus, which i don't think tells us a lot. so given this premise, why would you be interested in building more accurate postflop trees when solving for preflop?
additionally, while it is true that pio struggles to converge with multiple postflop sizes, it is simply a general thing that multiple postflop sizes don't affect average preflop evs much. ofc the abstraction loss is bigger for subtrees with larger sprs, but generally speaking if you use one of the unabstracted (nl) solvers and run simple trees vs complex trees over a lot of flops and average out the evs, there's not much of a difference to be seen. so if you have free ram for a sim, always allocate it towards reducing abstraction losses other than postflop bet sizes when solving for preflop.
I was mostly interested in hearing whether there was a reason for the AUTO tree builder to insert multiple sizings pretty much everywhere.
Fwiw, I have tested Monker's SB v BB equilibrium vs a benchmark Pio simulation (one geometrical bet size per street, 198 flops, 2bb/100 convergence) and found that {90 buckets, Perfect/Large/Large} and {30 buckets, Perfect/Perfect/Large} both reproduce the Pio results accurately. Frequencies same, and the shape of the ranges close to identical. In particular, the suited/offsuit polar regions in BB's 3B range are reproduced almost perfectly, which does not happen with default settings.
So we can verify that Monker reproduces accurate NLHE HU ranges when we push the accuracy settings up a bit.
Last edited by ZenFish; 09-04-2017 at 10:39 AM.