7 ms·
DeepStack: Expert-Level Artificial Intelligence In Heads-Up No-Limit Poker
- IMTDb 10y agoThis seems to be a major threat to online casinos. Is there any study on how they plan on fighting the AI's that will likely be flooding the market soon ?
- andr3w321 10y agoThe market has been flooded with them for years. They have various mechanisms for fighting them like mouse detection and scanning your computer for various programs and confiscating your funds if caught. In the end, people will always find a way around the detection methods though.
- SamBam 10y agoLike using a chess computer, it seems that there would be no way to protect against someone who relays the cards to a separate machine, and then enters the bids by hand on their own computer.
- Tenoke 10y agoActually, there are - most(all?) of the popular algorithms play in somewhat specific ways that are somewhat detectable (or at least flaggable).
- maaaats 10y agoSomewhat easy to detect a human using a computer, by comparing to what the computer would do, the rating of the player and the usual play of the player. However, a chess professional would be able to use a chess computer almost undetected, as they know what would look like "computer moves", but could use it to just choose the best of their own suggested moves.
- throw_away_777 10y agoThere are ways to detect cheating on chess websites, though these methods are not perfect. The two main signatures are: 1) Playing moves that the engine recommends consistently. This looks for people who are inhumanly good, even the top chess players in the world regularly deviate from engine suggestions. 2) Time between moves always takes a minimum threshold, the time needed to relay the position to the engine. Humans vary wildly in the time thought based on the position, in a way that cheaters do not. Neither of these detection methods are perfect, especially for small numbers of games. I don't think we will see large prizes for on-line chess tournaments because of possible cheating.
- Tenoke 10y agoWhen they released the initial paper a few months ago I looked into it and it sounds like this wont be profitable anywhere except the highest levels (although obviously the highest costs are in the training phase). This in combination with casinos already doing a lot for bot-detection makes me think that we are at least a few years away until this is an issue.
- corporateslave 10y agoI have a few friends in the poker community, bots have been playing extensively at all levels in heads up for atleast 5 years now. It can be assumed that this paper is way behind the state of the art.
- gdulli 10y agoHeads-up is a contrivedly simple scenario compared to solving poker in general. There have been bots for a long time that have co-existed with real players, and we're likely not close to a scenario where bots can reliably beat real players in general for real world scenarios. However, in theory even a break-even bot could be a gold mine because it can earn loyalty points. That's how it was the last time I was playing online, I don't know if loyalty programs still work the way they used to. Of those programs would just go away if abused enough.
- Tenoke 10y agoLoyalty points are paid out of (and lower than) the rake, so the bot still needs to be better than break even.
- alasdair_ 10y agoOnline poker has the concept of "bonus whoring" where one takes advantage of various loss-leader promotions offered by different sites. A typical promotion would be offering a 20% bonus on a $1000 deposit, with the caveat that you must play in, say, 1000 raked games before cashing out. It's possible to play in games where the total expected rake if the user breaks even is, say, $100, leaving $100 profit.
- backpropaganda 10y agoWhy do you say we're not close to solving Poker in real-world scenarios? The multiplayer version is likely to be harder, but is it so hard, that it'll take decades more?
- xapata 10y agoDepends what you're trying to achieve. If it's a bot that beats most humans, that probably already can be done. If it's a bot that earns more from a cash ring game than the world's best would against all levels of players... It took many years for a computer to bridge the gap between better at chess than almost all humans and better than the best human.
- andr3w321 10y agoAnyone know where to find the pseudo code? It references item 10 in the References and Notes which states See Supplementary Materials.
- foob 10y agoThey're included in the arXiv version of the paper: https://arxiv.org/abs/1701.01724 https://arxiv.org/abs/1701.01724
- PaulHoule 10y agoOne thing I haven't seem come up in discussions or media coverage about this is that the heads up game is entirely a different thing than playing with 5 other players.
- Tenoke 10y agoFor what is worth, heads-up is a very popular game, and from reading the paper, the approach should more or less be applicable to 6max or full ring, except it will be way more computationally expensive (but not unfeasibly so like before).
- jumpCastle 10y agoIn my understanding the approach is based on game theoretic principles that do not extend naturally to more than two players. If that's correct then it's not only a computational issue. See http://www.nature.com/news/how-rival-bots-battled-their-way-to-poker-supremacy-1.21580 http://www.nature.com/news/how-rival-bots-battled-their-way-...
- peternilson 10y agoYip, exactly this. The CFR algorithm they speak of is based on finding a nash equilibrium for each strategy-pair in the game tree. Something that I think can only be computed for 2 players, and no more.
- Tenoke 10y agoSurely you can adapt it to compute an approximation of a mixed Nash equilibria. My game theory isnt at a high level, but there are similar computations you can perform, and when I used to play poker, you'd have common calculations based on it for more than 2 players when you are playing with a limited stack. I dont see how that wouldnt be adaptable (abeit more computationally expensive) for more players, but I might be missing something.
- 10y ago
- maaaats 10y agoPretty helpful with the annotation along the way, explaining mbb/g and what it means, for instance.
- LeanderK 10y agoHow can one create visualisations like Figure 3 (assuming it's latex)[0]? I have to write my Bachelor thesis soon and need to improve my latex-skills. I only know to to use the math-mode and write text. [0] the paper: https://arxiv.org/pdf/1701.01724.pdf https://arxiv.org/pdf/1701.01724.pdf
- fhackenberger 10y agoProbably TikZ. There are several frontend UIs available for accelerating the process of creating diagrams. One is ktikz, which I initiated a while ago.
- freyr 10y agoYou can create images using your favorite image creation software (e.g. Adobe Illustrator, Inkscape, MS Paint, D3, anything) and embed them in a LaTeX document.
- nerdponx 10y agoDon't forget GraphViz. Although it's probably underpowered for that figure in particular.
- mikebjohanson 10y agoHi - I'm one of the authors on this paper. We had professional designers do the figures for this paper; Fig3 in particular was far better than what we could do ourselves.
- andr3w321 10y agoJust working my way through the pseudocode to try and gain a little insight. I think there may be a small typo. Line 29 VALUE should be VALUES?
- mastazi 10y agonon-pdf link: https://arxiv.org/abs/1701.01724 https://arxiv.org/abs/1701.01724
- grumpopotamus 10y agoAny hints on how to implement CFR-D on the GPU?
- mastazi 10y agoWith regards to Texas Hold'em, Conterfactual Regret Optimisation [1][2] was already able to beat humans in limit games, while it seems that this new approach is able to do the same in no-limit games. My naive understanding is that no-limit games are harder to crack because there are more possible outcomes, thus the available information (poker is an "imperfect information" game) becomes "even more imperfect". [1] https://arxiv.org/abs/1407.5042 https://arxiv.org/abs/1407.5042 (yes, the author is the same Oskari Tammelin who created Jeskola Buzz, a popular piece of music software). [2] http://www.nature.com/news/game-theorists-crack-poker-1.16683 http://www.nature.com/news/game-theorists-crack-poker-1.1668...