Is there any such discussion/analysis other than Doug Polk's $499 course on this exact subject?
I found a brief discussion by the human opponents of Libratus at:
http://www.pokerlistings.com/how-to-...-w/o-being-one
Starts 2/3 down the page
TL;DR below
I'm interested because I'm both a poker player and an AI (machine learning) researcher. I understand
Libratus' & AlphaGo's "Reinforcement Learning" approach vs. PokerSnowie's "Neural Network" approach.
TL;DR
-
Very balanced actions (similar hands with vastly different play)
- Bet-sizing was totally independent of pot-size (pot-size is just an arbitrary 'reference point' for human betting)
-
Pot overbetting is the 'norm' (even with, say, middle-pair). Is this 'in general' OR only for fold-equity purposes?
- Bluff & call frequencies
much higher than human players
- Bluffing and calling in situations where no human would ever consider such plays
- Libratus would always have a bluff available in any situation
- Libratus would 'tank' on the turn & river, exactly like a chess endgame (where there is always a provably correct play)