What happens when you run bots against other types/species of bots?

Does one type of bot figure out a way to win more somehow? Do they explode? What? Has there ever been a legitimate ‘bot off’ tournament or cash game online from which to extract undiluted bot v bot MDA that's worth reading?

My thick as **** presumption is that the most fundamentally precise bot would find a way to max exploit some tiny advantage in CPU or processing power and effectively print EV against all the other bots over infinite hands? (???) Or would electricity access win like in bitcoin? Would it boil down to who could counter adapt the fastest in between hands, predicting their opponent’s weaker adaptive capability first (and adopting strategies that exploit that comparative weakness the quickest before detection)?


Or will everything just equalise like a damp squib?

Bot end boss or nobody goes anywhere boss?

pass me whatever you're taking

I personally always look forward to posts by Ceres.

Pretty much what you see in your profile pic. One bot ends up riding the other one around like a little b****.

In all seriousness though they should battle the bots to see how they perform against each other over millions of hands.

There have been several bot vs bot competitions in the past. There used to be something called the Annual Computer Poker Competition, but it stopped running back in 2018.

The problem is that most adaptive poker bots - those able to adapt to any bet size and respond quickly - aren't public. So this makes it hard to find sparring partners. One of the few publically available bots is called SlumBot - this was the champion of the most recent ACPC. So most researchers use this as a benchmark to test their own privately developed bots.

We even used Slumbot to benchmark GTO Wizard AI

Here are the results of various poker bots vs Slumbot:



I think one of the Slumbot's biggest weaknesses is that it's bad at responding to sizes outside of its abstraction. Internally, it has a plan for specific bet and raise sizes. Then if it encounters something between those preset sizes, it basically just looks at the two nearest sizes and makes a weak guess as to what the new strategy should be. But that's just my uninformed guess lol.

----

GTO Wizard ran this 150k battle twice and were surprised to find that a dynamic 1-sizing strategy outperformed a more complex 6-size strategy.

Theoretically, a complex strategy should outperform a simple strategy, but the time constraints (7s per move) meant that a the simpler approach could reach higher accuracy.

GTO Wizard AI vs Slumbot - Complex Strategy
Win rate: 13.1 ± 4.2 bb/100



GTO Wizard AI vs Slumbot - Dynamic 1-Size Strategy
Win rate: 19.4 ± 4.1 bb/100




If you'd like to see a deepdive - Kevin Rabichow explored the match in this RIO video:
https://www.runitonce.com/poker-trai...se-vs-slumbot/

Woaaaah...19.4bb/100!?!

We have a winner


Thanks tombos, fascinating.

I can see why you guys bought Ruse now. No doubt the innards are top secret, but does this mean Nash as an absolute strategy is now relegated to second rate? And neural nets are ultimately required for true optimisation?

And presumably this thing/beast obliterates humans (fish/weak regs) for even higher WRs? Or is it outperforming bots because of inflexibility as you say, and would struggle to max out like this vs a competent adaptive reg?

(soz, probably answered in the RIO videos but I dont’ have an account)
.
Also interesting/amusing that minimising complexity also helps bots crush bots. +1 for simplification then.

yeah that’d do for me