Quote:
Originally Posted by Sam Ganzfried
I'm not sure how useful an exploitability number or lb would be. If I came to you in 2008 and said "hey, our bot won the LHE competition, and it has exploitability 250mb/g (while always folding has exploitability 750 mb/g)", you would've laughed in my face. But yet, that bot went on to beat top humans including Hoss_TBF and others with statistical significance in the man vs. machine competition, http://webdocs.cs.ualberta.ca/~games/poker/man-machine/.
I don't quite understand why you are linking exploitability to the question whether or not your bot would beat (top) human players, that was never my point.
It seems like most teams in the Computer Poker Competition follow the same approach to creating a NLHE HU bot: Find a good abstraction of the full game, find a strategy that is as close to equilibrium as possible within this abstraction, and finally map the strategy back onto the full game. The results of the competition can be interpreted as such that your bot (strategy) is beating all other strategies -- however, we have no measure of accessing how "good" or, loosely speaking, how close to GTO these strategies are in the first place. A lower bound of exploitabilty would give us some insight to this question and would also be a natural criterion to access the quality of your strategy.
From a different perspective, the article and some of your comments advertise that human players could learn a lot from looking at how the bot solves certain spots. But why would I as a human player spend time learning from a bot without knowing how good the bot is, or how much room for improvement there still is (especially when you are stating yourself that the strategy may look very differently when the algorithms fully converge)? Back in 2008, even though it seems clear that the -250mb/g bot was better than top human players at this time, I don't think it would've been the best strategy for anyone to learn from this bot, but rather wait until the bot applies a strategy that has a signifincatly reduced exploitatability, therefore giving the people trying to learn from it more confidence in the strategy.
It surely would be interesting to see human players playing vs the bot, but I think most of us are aware that at some point in time, NHLE HU bots will be better than most human players anyways, so there's no real point in proving that this is or isn't the case in 2014 already.
Quote:
Originally Posted by Sam Ganzfried
If I told you that our bot's strategy was super unbalanced on rivers, but in an arbitrary way on different rivers, how would you go about exploiting it?
Again, I don't really understand why you are aksing me this question as this seems unrelated to me +1ing a request for a lower bound of (theoretical) exploitability of your strategy. I'll answer it anyway: Since you just told me that any signal of the bot's (average) river strategy is arbitrary and therfore uninformative, I wouldn't apply a strategy that is trying to maximally exploit leaks in the bot's strategy on the river. However, I would approach river spots with the best balanced strategy my human mind can come up with. If your bot is really "super unbalanced" in river spots, my strategy is very likely to exploit the bot's strategy by default, but not maximally.