What do you guys think about Pluribus' (6max bot) style?
Posted by DonoDrapero
Posted by
DonoDrapero
posted in
High Stakes
What do you guys think about Pluribus' (6max bot) style?
The 6max bot Pluribus beat a group of professional players in a recent match. The hand histories are hard to navigate through, but some people were kind enough to paste interesting ones in the 2+2 thread ( https://forumserver.twoplustwo.com/29/news-views-gossip/first-ai-beat-multiplayer-poker-game-vs-top-pros-like-chris-furgeson-darren-elias-1747396/ ).
Since solvers have been around for a while (and therefore we can already check unexploitable strategies for HU) the most interesting aspect of a match like this to me is how an AI approaches multiway spots. And it's surprising, because based on the hands it seems to be wildly aggressive. E.g. opens CO w/ K7dd, BTN calls, BB calls. Flop is 854r, bot cbets 2.3x pot. Is there an assumption BB should be leading most of his strong range? My thinking was overbets are not used multiway, because a single players pot share rarely allows him to.
Another example - 4way flop, bot shoves over cbet w/ 65dd on T42dd, SPR 10. I can sort of understand this one, but it's still interesting to think this is the highest EV line according to the bot.
Loading 39 Comments...
I wrote a program which parses the histories although its quite rough. I'm sure someone will do a better job and they will all get uploaded but in the mean time for people who want to pick through some of them right away, here are sessions 100-100b and 40-40b
https://pastebin.com/cG5pdgvz
https://pastebin.com/0ppR6Ybh
Happy to discuss interesting hands, and willing to upload some more histories later if they don't appear first.
Nice! It's pretty readable. A format that could be imported in trackers would be much better though.
Mr Pink seems like the only good reg based on reading first 20 hands. Linus?
I'd consider modifying the program to produce importable histories if the format isn't too complicated. If someone wants to PM me some examples of supported formats, I'll give it a try, but I do think it will be somewhat of an iterative process to get it just right.
This has been my assumption aswell, really interesting that a bot would choose to do overbetting. An obvious explanation would be that the overbet is used at a very low frequency with very specific range. I'm curious about this too.
Hand 26
MrPink[9c5c] folds
Pluribus[9sAs] raises 100
Eddie[8dKs] folds
MrBrown[4h9h] folds
MrBlue[3s4s] calls
Bill[7sAc] calls
Flop: Kd6cJc
MrBlue[3s4s] checks
Bill[7sAc] checks
Pluribus[9sAs] bets 300
MrBlue[3s4s] folds
Bill[7sAc] folds
So I miss read HH here and it's just a bet of 300 into 600 for 50% pot.
Odd call down at surface level but probably is some sort of unblocking effects or something.
Hand 36
MrBlue[7d2s] folds
Bill[3hKs] folds
MrPink[8h4s] folds
Pluribus[3s3c] raises 125
Eddie[AcJd] raises 875
MrBrown[JcTh] folds
Pluribus[3s3c] calls
Flop: 6d5d6s
Eddie[AcJd] bets 690
Pluribus[3s3c] calls
Turn: Kd
Eddie[AcJd] bets 1230
Pluribus[3s3c] calls
River: 9s
Eddie[AcJd] bets 6980
Pluribus[3s3c] calls
I tried to examine the hand but I don't understand turn call and river call doesn't look great either. Might of course be missing something.
I assume you ran this through a solver. Did you get the EV of calling? If it is only a little worse than folding the solver won't do it at all, but the cost of such errors will be low and perhaps easily overcome by more significantly 'superhuman' play elsewhere.
Or maybe the bot just isn't that good. This post expresses some skepticism: https://forumserver.twoplustwo.com/showpost.php?p=55267005&postcount=66
I've ran a sim on this one as well, both turn and river calls are losing on my sim
Yes, solver but not sure if I got everything correct. Only have my laptop without preflop solver while travelling during the summer.
Turn call looks straight up bad if I got bet size correctly.
Interesting play by MrPink whom I believe to be Linus:
Hand 53
MrBrown[5h9h] folds
MrBlue[Qd2s] folds
Bill[4hKc] folds
MrPink[Ks8h] raises 150
Pluribus[7c3c] folds
Eddie[TcAd] calls
Flop: 3hTd8c
Eddie[TcAd] checks
MrPink[Ks8h] bets 750
Eddie[TcAd] calls
Turn: Ah
Eddie[TcAd] checks
MrPink[Ks8h] checks
River: 6d
Eddie[TcAd] bets 3500
MrPink[Ks8h] calls
My guess for hand selection is that 8h blocks 88 set and 1 of the 2 common 2pair combos that's frequently in BB's range (T8s). Probably more of a balance play to be able to improve on 8 pairing cards. Gonna look at this later with solver.
Why do you think this is Linus/where can we find information on who the human players were?
MrPink seems to play like a world class player.
are all hands played at 100bb's deep?
I heard someone mention rake free, no ante and 100bb but I don't have a good source. There are some HHs that seems to go all in with 10k chips at 50/100 blinds too. They also usually play these bot challanges with fixed stack depth so the bot hasn't solved full 6max but just exact stack size. It should be able to solve arbitrary stack depths though.
Pretty sure there is no rake and no ante. The original files include a win/loss for each player and they add up to zero (and are zero for non-blind hands that don't play). The blinds are 50-100. The paper mentions 10000 as the maximum allowable bet so I assumed 100bb. I am not sure if the stacks get reset each hand.
I think raise in HHs is raise over last amount and not raise too or something. Not sure. I mean 110 raise is raise by 110 to a total of 220.
In the histories I've posted that is correct. In the original log files, the raise ("r") number is chips put in since the beginning of the hand (all streets).
Thanks for posting this. Would be awesome to get those HH readable on a tracker as Kalupso suggested!
PM me some sample histories in an acceptable format and I will look at modifying the program. I haven't played online in years and I have no current history files.
https://forumserver.twoplustwo.com/showpost.php?p=55276019&postcount=132
Hey guys looks like someone uploaded the hands imported on PT4 to youtube.
Part 1
Part 2
Part 3 with hands where the AI donks
It seems some of those leads are objectively bad, no?
E.g. BTN opens, SB flats, BB (AI) calls w/ 67o. Flop is AJ6r, AI leads 2x pot.
he plays interestingly pre( that big 4bet sizing) and postflop leads/3bets.
Usually big sizings pre are just constraints to the abstraction
but overall very solid in multiways, after leading etc. Obviously no need to mention SRP where he crushes.
Intresting mw hand, Linus and Pluribus in action
MrBlonde (UTG): §10,000 (100 bb)
MrWhite (MP): §10,000 (100 bb)
MrPink (CO): §10,000 (100 bb)
MrOrange (BU): §10,000 (100 bb)
Pluribus (SB): §10,000 (100 bb)
MrBlue (BB): §10,000 (100 bb)
Pre-Flop: (§150) Hero (MrOrange) is BTN with 6♣ 6♥
2 players fold, MrPink (CO) raises to §210, MrOrange (BU) calls §210, Pluribus (SB) calls §160, MrBlue (BB) calls §110
Flop: (§840) A♦ J♦ K♦ (4 players)
Pluribus (SB) checks, MrBlue (BB) checks, MrPink (CO) checks, MrOrange (BU) checks
Turn: (§840) 7♦ (4 players)
Pluribus (SB) checks, MrBlue (BB) checks, MrPink (CO) bets §840, MrOrange (BU) folds, Pluribus (SB) calls §840, MrBlue (BB) folds
River: (§2,520) (2 players, 0 all-in)
Total pot: §2,520 (Rake: §0)
Showdown:
MrPink (CO) mucks 7♣ J♣ (two pair, Jacks and Sevens)
(Equity - Pre-Flop: 51%, Flop: 58%, Turn: 18%, River: 18%)
Pluribus (SB) mucks 2♣ 2♦ (a flush, Ace high)
(Equity - Pre-Flop: 49%, Flop: 42%, Turn: 82%, River: 82%)
MrPink (CO) wins §2,520
River off suit K:
Linus shove and Pluribus fold.
How do you know Linus is Mr. Pink?
I have seen lots of hands by Linus so Im pretty sure its him. Other human players seems bad, from what ive seen so far
hint - his pokerstars avatar pink color too
BTN open 2.5x, SB call, Pluribus on BB call.
Flop: 6s Js Ad, Pluribus lead 2x pot with 67o bdfd.. o.o
Importable hands
Great job ethanrox, and thanks for liberally sharing it!
Can somebody enlighten me - I am not 100% on track. It read that Pluribus "beat" the opponents and the graph they presented, showed a winrate of close to 5bb/100. The import shows a loss.
Furthermore, it says, that 5 bots played against one human, each human played 5k hands. Why does the graph they present shows 10k hands - which is exactly the number that is contained in the import file? Is that an arbitrary selection? Likely I just missed the information in the mass, so gratefully for everyone who points me into the right direction. :)
BigFiszh
Credit should go to Kevin Wang who made the conversion.
I glanced at the scientific paper because of the mismatch between results and ... results -> they have found a method AIVAT that can give you your "true winrate" based on a smaller sample but with a certain margin of error. So the true winrate is something like 4.8bb/100 when they have used this special methodology but within a margin of error of 1.3bb or something (cant remember exact numbers).
Pluribus style is as important as Deepstack and Libratus were before. It will change the game. Detailed analysis is worth it. Looking at 10K hands from database, we find a lot, I only looked at part of preflop for now, but see :
when folded to SB, 50/50 limp/bet from every hand in the range (roughly). Variations of bet size preflop, mostly 2.0, 2.25, 2.5 . Strange 3bet hands in position, like 55.
I put my findings as they come on my pluribus blog, maybe you like it.
Just looked at Pluribus cbet frequencies and sizings. When he cbets (less than most) 10% of cbets are 200% pot so yes a lot of over bets. https://pluribus-poker-ai.com/2019/08/15/pluribus-poker-flop-cbet-strategy/
Added all the flop overbets by Pluribus here: https://pluribus-poker-ai.com/2019/08/23/pluribus-overbets-on-the-flop-poker-ai/. Bear in mind the model limited bet sizes to only 50%, 100% and 200% for single raised pots. Ie Pluribus never ever tried other sizings.
Be the first to add a comment