Quote:
Originally Posted by daimonos
For example- button opens, I 3bet 44 in SB, BB coldcalls, button folds. Snowie says check. I bet half pot anyway because I'm stubborn. Snowie tells me that full pot is the optimal bet size, despite the fact that the 1/2 pot bet has an EV of 0.85, while the pot size bet is -0.04. What gives?
This is due to the slightly odd and confusing way the artificial neural net works, and how the "bot" was programmed to choose just one ('best overall on average') bet-size for its entire range in a particular spot, rather than "range-splitting" and balancing multiple ranges and sizes at once.
That is to say, the network might have learned that hand XX makes 0.85bb when it bets half pot with that exact combo, but it thinks it maximizes the EV of its entire range/strategy by betting pot with its range, and since that specific combo will lose money by betting pot, it checks with it instead.
As a quick example, it might be the case that betting 2x pot would make the most money with the nuts, but the range as a whole does better by betting 1/4 pot, because there are lots of hands that don't make money as overbets, but can make money by betting small. In order to be able to profitably bet
lots of hands for 1/4 pot, Snowie also has to choose the small sizing with the nuts, or otherwise it would be too exploitable/face up. In the example you found, it probably decided that its range wanted to bet big (pot) because it could polarize more often, and utilize additional fold equity with its bluffs, but the combo in question didn't belong in that polarized range, because it was neither a fat value bet, or an obvious bluff. You'd like to bet smallish with that hand, but the best strategy
for the range as a whole is to bet big.
With other solver software it can be found that an even higher overall EV can be gained by betting small with some parts of a range, and large with other parts (with some combos using both sizes), but Snowie was programmed to only use one size at a given time.