Its clear that for holdem its better to use a 7-card evaluator since you can treat your hand as a 7 card hand as opposed to a 5 card hand. But for PLO I don't see how you can do this.

Is it better to use a five card evaluator, like Cactus kev, and evaluate every five card combo (maybe sort some of the combos out b some process) or is it better to do what poker eval library seems to do, which is to evaluate the hand as a holdem hand and then check if the right number of cards from deck and hand are present. It says in the poker eval source code that Omaha evaluation is expected to run 60 times slower than holdem eval (iirc). If thats true can it really still be faster?

Would be glad if someones whos tested this would reply.

Mvh
Inga

Yeah, also interested on this.

The only solution I see is the "slow one". I think this is what poker eval does.


Treat each player as having 6x 7 card hands.

With 4 holecards {a,b,c,d} and player must use exactly two each player has 6 combinations of hands: {a,b} {a,c} {a,d} {b,c} {b,d} {c,d}

Call the board z1-z5

So it is
{a,b,z1,z2,z3,z4,z5}
{a,c,z1,z2,z3,z4,z5}
....
{c,d,z1,z2,z3,z4,z5}


So for each player evaluate their 6x 7card hands. For full ring game you are talking about 60 hands. The best hand out of the 60 is winner.

The only thing that is confusing to me is I can't see why it would be 60x slower. It is ranking 60 hands vs ranking 10. That can't be 60x slower can it?

If poker eval is really is 60x slower then I imagine it is doing something less efficient and likely you could speed it up using algorithm similar to the above.

With a five card evaluator, for a heads up scenario and 600000 iterations montecarlo search you would have to evaluate 1,2 million hands I think. With 60 Cactus Kev evals per hand you should get 72 million Cactus kev evals total I think.

That takes awhile, would you think it likely to be faster with poker eval despite it being 60 times slower than holdemeval?

Mvh
Inga

This sounds logical but if I recall correctly the twoplustwo evaluator uses a lookup table that doesn't require card sequences. Could it not be improved then by placing the board cards before the hole cards (i.e. {z1,z2,z3,z4,z5,c,d}) and 'bookmarking' the position in the lookup sequence at the conclusion of the board cards. The final iteration of 6 combinations of hole cards would loop for only the final two steps. That would have to be faster than doing 6 full 7 card lookups/evaluations.

Looking at the source code for evaluating omaha hands in poker eval, it looks like it does a huge amount of extra work.

First it creates every legal combo of hole cards and stores in an array, then same for board cards. It then loops through every five card combo of hole and board cards, so 60 iterations. So effectively it does a five card evaluation.

The dividing the holecards into separate masks and same for board cards must create a great deal of overhead. Sort of doubt this could be faster than Cactus given the extra work. Will test it when I have some time left over.

I know this thread is old but I have made two evaluators for Omaha inspired by the 2+2 sources for TH, one of them is table driven and evaluates +41 million hands a second (using 8 threads on a very old i7 @ 3ghz at the cost of 512mb of ram)...