Quote:
Honestly, we had just assumed computing exploitability is hard
So you guys are using an algorithm to approximate something you can't measure?
Quote:
, since we use imperfect-recall abstractions
Calculate ranges at the end of every line (up to turn), you can do that cause you play flops/turns fast.
Solve rivers with those ranges (or recreate them in any other way are doing during play but just solving them with assumed ranges is the most obvious way), profit. You can use even 5 bet sizes on the river, it will be fast anyway, especially on the kind of hardware you have.
As it's depth-first-search you don't need much memory for that either.
Quote:
It's not clear to me if it's doable, and what the scientific value is of computing a bound on our exploitability like this.
Is there any scientific value in approximating something you don't have a measure for? I mean the results can be w/e and you will not know. If you can't assess it you can't experiment, you can't say if a tweak is good or not etc.
Quote:
chess bots are better than the bests, isnt it still a great game?
Chess is way harder than poker though and top programs are nowhere near optimal play yet (although way stronger than humans). The problem with Holdem is that it's possible to solve it or at least approximate it to make it essentially solved (like Alberta team did with limit). It will happen to HU NLHE sooner than solution to chess happens.
Good thing though is that multiplayer games are unlikely to be solved and equilibrium in those isn't as valuable anyway.
Last edited by punter11235; 05-07-2015 at 02:35 AM.