Quote:
I understood about half of what you just said but good to know.
Let me reword it then because the tradition of naming stuff and using cryptic language to describe simple concepts is better left to academics and their papers.
Quote:
Some older solvers/AIs couldn't take blocker effect into account because they used heavy card abstraction (bucketing into categories etc.).
Card abstractions are ways of grouping similar (or kinda similar) hands together in hopes to reduce the size of the game and deal with only few categories (buckets) instead of all 1326 combos. The advantage is that you use less memory and you can compute things faster, disadvantage is that you lose the information about specific cards and can't tell a difference between say AsQh and AdQd on 7s 5s 3c 8s 3h board.
Quote:
From not very descriptive overview of Tartanian7 I've seen it seems the blockers should be taken into account (because they run MCFR without bucketting).
CFR is one popular self-play algorithm ("it plays against itself and evolves over million of iterations" in marketting terms), it just means it adjusts strategies trying to exploit at every step hoping that after enough steps of that for both players the almost equilibrium is reached (aka GTO).
MCFR is the same thing but instead of considering all the possiiblities (like all turn/river cards and all possible holdings in player's range) in all iterations it just plays some hands vs itself and tries to adjust from that. MCFR is exactly what human players do.
The description of Tartanian7 suggested that the way that particular bot does it is just play one hand at a time vs itself and adjusts. There is a way to make it mathematically sound if you play enough of those. If you only sample one hand for both opponents at every iteration then the blocker effect will be taken in consideration (because it will be clear over long run that AsQh has different ev than AdQd on that spade heavy board).