Quote:
- PioSolver cannot figure out that a given strategy should be used 0% of the time. It will decrease from 1% to .5% to .25% etc. but it will never reach zero. In the meantime, the EV with the extra strategic choice is not as good as the EV without it.
That the solver is slow to reach 0 is a known issue which is hard to fix without making an average case worse (or without using significantly more memory). It shouldn't be a problem with pure preflop solutions though as you can just use postflop solver for that and it they should reach 0 comfortably.
Quote:
In the meantime, the EV with the extra strategic choice is not as good as the EV without it.
Can you paste the config? I would be very surprised if it's the case with any noticeable difference in rakeless trees.
Quote:
- This effect is the same when several bet sizings are available; often one sizing (sometimes two) will prevail, however the % for the other sizings, although very small, do not reach zero. Yet EV is not improved with a wider choice of sizings.
The trees are big and it takes a while, again it's important if you use rake there. If you do, there are other reason for that, if you don't the EVs should very quickly be accurate and you shouldn't notice it.
Quote:
- Having the possibility to setup a starting point for solving would drastically improve the search for better strategic options, e.g. if PioSolver computes 73% push 27% fold for the first player, then we should be able to start 73% push 0% limp 27% fold the new strategy. If the new strategy drastically improves the EV, then the gradient (or whatever you use here) will tell the solver to increase the % of limp. Some "Build Tree using current solution as a starting point" button would be awesome.
It's complicated and not easy. If it was that easy it would already be available.
I can make the algorithm to get to 0 quickly and reliably in those cases. The problem is that it's not as good in average case then.
We would of course like to improve the algorithm but a quick fix like that won't help.
Quote:
Maybe I missed a link where you explain how these were built / how the balance between simple/exhaustive/accurate is created?
We have used the results computed on all 1755 flops to find subsets which approximate the EVs most accurately. More here:
http://piosolver.myshopify.com/blogs...the-whole-game
Quote:
I really would like to be able to use in PioSolver the complete 1755 flop subset with the exact weights that describe perfectly the 22100 flops.
If you have enough RAM or trees simple enough you can do it. You can find the full subset here:
http://pastebin.com/rJKh4sxN
There is a bug in currently shipped version of the solver which will cause a crash if you use ~1500+ flops. Please email me with your key (or registered email) so I will ship a bugfix for you.