Could somebody please explain why every hand is reset to 200 BB?

Is this because it's the only correct way to test the difference between humans and AI? Is the AI unable to adjust to differing stack sizes as would be seen in a normal heads-up match? (Because it's too complicated/complex for the machine?)

It appears to me that this negates a lot of the knowledge a human expert possesses :/ but then again, maybe I'm just clueless!

I wonder if this is one of those cases where Rich's original post was serious, then he realized that he's ******ed, so now he starts leveling to cover up his initial **** up

The bot is trained to solve with 200bb stacks. It also helps to make the duplicate format relevant.

It's a convention. From https://arxiv.org/pdf/1701.01724.pdf (Deepstack working paper, p15):

And of course varying stack sizes make the game more complex so that convention definitely works in favor of the bot(s).

Was it announced if every hand is in a vacuum or the bot has a memory of previous hands vs different villains?

I always thought that each hand in a vacuum was a really unfair way to gauge the bot's ability. Online bots remember every single detail of your play which is what gives them that extra edge.

It does not adjust throughout and to different opponents. So no memory.

The point of this bot is to be non-adjusting approaching GTO, which is a non-adjusting strategy. Optimal is optimal, it doesn't change or react to what you throw at it.

GTO gives no ****s how you play.

Didn't Doug say the Claudico team was in the lab making adjustments after each day of play? Why would it be different now?

This is how it should be, however a) as mrno stated above, they did perform overnight changes to Claudico last time (I believe mainly to counter the excessive minbetting from the humans) and b), as I quoted above from the riverscasino site,

there seems to be something similiar going on this time, but since no real info is published on how Libratus works, we can't know for sure.

The gametree is way too big for it to play GTO anyways.

Do you mean in the 200BB 'reset' format they're playing currently or in general HUNL -or both/either any HU format?

Quoted from above:

Libratus’s strategy is not based on the experience of expert human players, so its game play could differ markedly from the pros. It uses algorithms to analyze the rules of poker and set its own strategy, based on approximately 15 million core hours of computation at the Pittsburgh Supercomputing Center (PSC). Libratus will continuously sharpen its strategy during the Brains vs. AI competition, performing computations with the PSC’s Bridges computer each night while the pros get some shuteye. During games, Bridges will perform live computations to aid Libratus with its end-game play.

So it seems to be saying that extra-Libratus calculations do affect the play in some unknown way.

I can guess that since they have super computer hours rented they use the night to keep the bot running and updating (bettering) its strategy. It still doesn't adjust to what happened previously though

I don't know why you say that. I am not disagreeing, but that is not how I interpret the above quote.

The smaller the stacksize the smaller the gametree. One would expect hyperturbo HU to fall first to the computer. 200bbs is quite deep and honestly the computer beating humans there would be quite scary. I don't know if there is an estimation of the gametreecomplexity for poker, but my guess would be it's more complex than chess, so we probably will never see "perfect play".

But it obviously doesn't have to play perfectly to beat humans.

edit:

http://poker.cs.ualberta.ca/publicat...rt-nl-size.pdf

this seems to have some information about the gametree.

It also obviously requires a different approach from chess.

chess is actually super super simple compared to nlh (200bb deep or otherwise)

Not looking good for humanity.

5% of hands. 13 buyins. still a bit early, although it might just be a sign of things to come

I still have hope for humans


129000/20000 = 6.45 buyins not 13 (200bb buyins though i guess to be fair)

All of them being down is a very bad sign.

If humans lose this there will not be a rematch against "better" human HU players.

I am very bad at maths, but isn't that 21bb/100 so far?

Following this with interest... Would someone mind clarifying - are they each playing a random selection of hands as dealt by the dealer? Or will all the pros be dealt the same cards/situations as each other over the course of the challenge?

Was it ever revealed what the bot had when Dong folded the T9 straight?