Quote:
Originally Posted by ArtyMcFly
One of the problems is that introducing more bet-sizes increases the size of the decision tree - and the EV results database - which tends towards infinity. A neural net like Snowie could learn how to play a game that only used pot-sized bets, but if the half-pot sizing is introduced as an option, all the previous learning is rendered null and void, and a new "solution" needs to be computed that allows both sizes to be used at the optimal frequencies. From there, if you add another bet-size option, the solution needs to be re-computed once more, making it even more complicated (it's essentially a whole new game with another layer of options). And then you need to solve the river play for spots where the turn was played incorrectly at some (arbitrary?) frequency.
I used to agree with every word here. However, now I have doubts about the complexity of alternate bet sizes.
I am really just getting started with multiple street no limit poker math, but here is my take on it thus far.
Please correct me and help me learn if any of this is way off track!
If stacks are infinite, there should be an optimal bet/raise size on every street. It is somewhere around the size of the pot. In toy poker games it is exactly the size of the pot. This is because bets of other sizes unnecessarily move the indifference point boundaries within range and lead to less range economy. Smaller or larger bet sizes also lead to smaller game trees, not larger game trees.
Lets look at the overly simplified toy game where only one player can bet and this bet can only be called or folded. Real poker has tons of options, and more options creates more indifference points in the ranges, but the overall range portion that survives each round of betting should be similar to the toy game.
A bet that is full pot brings along the most combos
per dollar from both ranges. This is because all bets, regardless of size, must reside inside the bettors balanced range, and non-optimal smaller or larger decrease the overall payoff to the bettor. Smaller bets sacrifice too many bluff combos and larger bets sacrifice too many value combos.
Furthermore, when a bet can be raised, the non-optimal size of the initial bet can be exploited, since the initial bet must be balanced. Balanced bets must come from portions of range that are bounded by indifference points. When a bet size is not optimal it comes from a skewed portion of range and can be raised (or folded) for pure profit by an opponent's hand that is outside this skewed range.
Optimal bet sizing reduces the chance of a bet being raised, such that a raise must in turn also be optimal and can not be pure profit.
Lets look at a bet of twice the pot on the river.
To be balanced, such a bet will come from two different locations in a bettors
surviving range. There will be a value to bluff ratio of 40% bluffs and 60% value. The caller knows this and knows that such a bet comes from the top 60 percent of a bettors value range, and the bottom 40 percent of the bettors bluff range.
Now play the hand backwards with 2x pot bets on each street.
The bettor has 56 percent bluff and 44 percent value on the turn.
The bettor has 78.4 percent bluff and only 21.6 percent value on the flop.
The bettor has run out of value!
Now lets look at a half pot bet, again only considering calling each street. Now the bettor has a ratio of 75 percent value and 25 percent bluffs on the river.
Now play the hand backwards with 1/2 pot bets each street.
The bettor has 31.25 percent bluff and 68.75 percent value on the turn.
The bettor has 39 percent bluff and 61 percent value on the flop.
The bettor has too few bluffs!
Now lets look at a pot sized bet. The bettor has 66 percent value and 33 percent bluff on the river. Now play the hand backwards with pot sized bets each street.
The bettor has 44.44 percent bluff and 55.56 percent value on the turn.
The bettor has 59.25 percent bluff and 40.75 percent value on the flop.
Now the bettor will get the most use out of both ranges. This ensures that the bettor will profit from over half the bluff range but still not run out of value by the river.
Now, if all these scenarios are the result of preflop play that has the options of bets and raises, then we can see how a player who arrives at a flop could really run out of bluffs or value by the river, if a healthy portion of bluffs or value was already reduced by preflop betting.
It should be pointed out that this toy game definitely is rigged. I believe, not for sure, but I am guessing that if a bet is required on the river, the caller has the advantage. If the bettor can instead checkback the river as well as bet, then the game favors the bettor.
If I recall what the humans said about Libratus, it was that Libratus could overbet the pot freely, but would then find enough bluffs by bluffing bottom value combos (hands with showdown value). This would be required by the above toy game.
Last edited by robert_utk; 11-20-2017 at 03:51 AM.