How easy would it be to modify this code for other poker game types? about open-pure-cfr

[rggibson]

Hey there,

Yes, the code could definitely be modified to omaha and low hand games.

The hand evaluation happens in acpc-server-code/evalHandTables, though it is pretty hard to digest and I believe it is hold'em specific (I didn't write it). However, for pot limit omaha, you could use the hold'em evaluator as a black box and simply iterate over all 6 pairs of 2 cards from the 4 hole cards and take the max.

As for suit symmetries, they actually aren't handled at all in the current code. The best way to integrate them would be to create a new class derived from CardAbstraction that assigns the same bucket number to hands that only differ by suits. Check out card_abstraction.*pp for that stuff.

Hope that helps!

[githubToolTester]

Very helpful, thanks!

Alex

On Wed, Oct 19, 2016 at 1:35 PM, Richard Gibson [email protected]
wrote:

Hey there,

Yes, the code could definitely be modified to omaha and low hand games.

The hand evaluation happens in acpc-server-code/evalHandTables, though it
is pretty hard to digest and I believe it is hold'em specific (I didn't
write it). However, for pot limit omaha, you could use the hold'em
evaluator as a black box and simply iterate over all 6 pairs of 2 cards
from the 4 hole cards and take the max.

As for suit symmetries, they actually aren't handled at all in the current
code. The best way to integrate them would be to create a new class derived
from CardAbstraction that assigns the same bucket number to hands that only
differ by suits. Check out card_abstraction.*pp for that stuff.

Hope that helps!

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#3 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AEuByJrfT5Y6jD1kavkYMvjijfPDrnDpks5q1n8dgaJpZM4KbXb4
.

[githubToolTester]

One other question, have you in any way compared the performance of this
code base to commercial hold'em solvers (SimplePostflop / Piosolver /
GTORangeBuilder?)

Also just to make sure I am not completely misunderstanding, the Pure CFR
algorithm samples from pure strategies but still produces a mixed strategy
as its output, correct?

Thanks!

On Wed, Oct 19, 2016 at 5:50 PM, Alex Sutherland [email protected]
wrote:

Very helpful, thanks!

Alex

On Wed, Oct 19, 2016 at 1:35 PM, Richard Gibson [email protected]
wrote:

Hey there,

Yes, the code could definitely be modified to omaha and low hand games.

The hand evaluation happens in acpc-server-code/evalHandTables, though
it is pretty hard to digest and I believe it is hold'em specific (I didn't
write it). However, for pot limit omaha, you could use the hold'em
evaluator as a black box and simply iterate over all 6 pairs of 2 cards
from the 4 hole cards and take the max.

As for suit symmetries, they actually aren't handled at all in the
current code. The best way to integrate them would be to create a new class
derived from CardAbstraction that assigns the same bucket number to hands
that only differ by suits. Check out card_abstraction.*pp for that stuff.

Hope that helps!

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#3 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AEuByJrfT5Y6jD1kavkYMvjijfPDrnDpks5q1n8dgaJpZM4KbXb4
.

[rggibson]

I have not compared this in any way to commercial hold'em solvers. I've never actually played with them, but from my understanding, they only solve games after the flop. Pure CFR solves the entire game (though it could be altered to solve for postflop strategies only given static preflop ranges).

Correct, pure CFR samples pure strategies but still produces a behavioral strategy with mixed action probabilities at each information set. Like all CFR variants, it converges to a Nash equilibrium, which for many games is necessarily mixed.

[githubToolTester]

Great thanks so much for the quick reply!

If you are curious both SimplePostflop and Piosolver solve starting from
both postflop with static starting ranges and from preflop. They do require
a specific bet sizing abstraction but they do not do any hand / board
abstractions aside from strategically identical suit symmetries.

On Wed, Oct 19, 2016 at 7:16 PM, Richard Gibson [email protected]
wrote:

I have not compared this in any way to commercial hold'em solvers. I've
never actually played with them, but from my understanding, they only solve
games after the flop. Pure CFR solves the entire game (though it could be
altered to solve for postflop strategies only given static preflop ranges).

Correct, pure CFR samples pure strategies but still produces a behavioral
strategy with mixed action probabilities at each information set. Like all
CFR variants, it converges to a Nash equilibrium, which for many games is
necessarily mixed.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#3 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AEuByGyXucurtK8cOd1dIP7W0fE4OqOPks5q1s8NgaJpZM4KbXb4
.