Quote:
any updates on the preflop solver, will we be seeing it in the next few weeks?
Yes. I was about to make an announcement so I can just as well make it answering this post.
News about preflop solver:
-first release is going to be next weekend
-the solver works locally, in RAM (for now)
-it works on subset of flops which you can choose
-
minimum requirements: 32GB of RAM, requirements for comfortable usage 64GB of RAM
-it works on cloud instances/dedicated servers as well which are likely to be the best options to run it;
-because of many memory optimizations as well as good flop subset selection (subset of various sizes will be added to the solver) the solutions are close to perfect even on very small subsets; for example I've run it for 2 hours on i7 3770k with 32GB of RAM for 7bb preflop game which is became a benchmark of sorts (because Will Tipton used it in his books and it was the only publicly available semi-solution you can compare against so naturally other people solving preflop have chosen it) and I've got the following results:
1)SB opens:
http://i.imgur.com/SXc4ae0.png
2)BB responds to a shove:
http://i.imgur.com/R7i5FVx.png
3)SB limps and responds to a minraise:
http://i.imgur.com/ABdkrlR.png (notice that the viewer shows the raise size as 60 instead of 70 but this is a glitch in a viewer and the solution is for 70$ (7bb) all-in)
4)SB limps and responds to a shove:
http://i.imgur.com/7hSVryD.png
The evs (5$/10$ blinds):
EV OOP: -0.291
EV IP: 0.291
MES OOP: -0.270
MES IP: 0.309
Exploitable for: 0.020
So SB wins 0.029bb/hand
If you compare this solution to recently published solution on all 1755 flops you will notice that our solution on a subset is way better than the one Tipton got and is almost perfect combo-by-combo match and exact EV match even though it was run on a computer with 32GB of RAM (and using 25GB). The reason for results being as good are:
-memory optimizations, so we can keep more flops in RAM; it was run on 58 flop subset and took 25GB of RAM, it means the whole game would be just shy of 800GB;
-good flop subset selection (check out the blog post about it, the flops were improved since then and we have subsets of every size available)
-no postflop abstractions so EVs from them are exact
We tested it on 100bb games solved before by other people and again it's almost perfect combo-by-combo match.
It means you will be able to experiment with preflop games and solve them using slight variations in bet sizings/locked ranges on a good home desktop (with 64GB of RAM and even toy with it with 32GB as I have) or on one cloud instance/dedicated server - no need for distributed solver and huge bills (although if there is demand we will likely provide it)
Pricing:
-The preflop solver is going to be a part of PioSOLVER-edge version; one activation of it is added there (and normal activations will be reduced to 2 to keep total at 3; it of course doesn't influence people who already bought the edge version or buy it before the announcement (they get 3 normal ones + 1 preflop solver))
-Additional activations will be purchased separately and the cost will be around 500$/activation
This means that for 1100$ and a not so big electricity bill you will be able to solve NLHE preflop in your home.
More details next weekend with the release. I am happy to answer not too specific questions for now.
Last edited by punter11235; 12-07-2015 at 06:56 AM.