noambrown

It's in the supplementary material of the 2019 paper: http://www.cs.cmu.edu/~noamb/papers/19-Science-Superhuman_Su... . Look at the "Variance reduction via AIVAT" section.

The four humans were getting $120,000 between them. Their share of that was dependent on how much better they did than the other humans. That means there was no incentive to collude.

Top pro poker players understand the value of money. They weren't treating it as a freeroll and anyone that has seen the hand histories can confirm that.

That's not true in practice for poker. Pluribus showed that if you run CFR in multiplayer poker you get a solution that works great in practice. Multiple equilibria are certainly a theoretical issue for many games, but poker conveniently isn't one of them.

Bots are superhuman in self-play Hanabi: https://ai.facebook.com/blog/building-ai-that-can-master-com...

The remaining challenge is getting it to play well with human partners. Doing that requires modeling human conventions rather than learning weird bot conventions. That's hard because while you can collect essentially unlimited data through self play, it's hard to collect a lot of data playing with humans using reinforcement learning. AI algorithms are really bad at sample efficiency.

Normally 10,000 hands would be too small a sample size but we used variance-reduction techniques to reduce the luck factor. Think things like all-in EV but much more powerful. It's described in the paper.

Bots are superhuman in no-limit Texas hold'em. Libratus beat top humans in two-player in 2017 and Pluribus beat top humans in six-player in 2019: https://www.science.org/doi/abs/10.1126/science.aao1733 https://www.science.org/doi/abs/10.1126/science.aay2400

It's shocking that the reporter didn't mention these results or anything else more recent than 2015.

Bridge has a similar challenge, though from what I understand Bridge AIs are not superhuman yet. I suspect our techniques could be applied to Bridge, though they may need to be adapted a bit.

The imperfect information in Hanabi absolutely matters a ton. It's not an interesting game without it.

Thanks! We're looking in a few different directions, but one thing I'm excited about is mixed cooperative/competitive settings. In poker, there is no room for cooperation. In Hanabi, you are 100% cooperating with your teammates. But most real-world situations, like a negotiation, are somewhere in between. The AI techniques for these settings are not too strong yet.

Open source it, learn from it, and build upon it to continue to push forward the frontier of AI.

Definitely!

In terms of Hanabi, this bot arrived at conventions that are pretty different from how humans play the game. We invited an advanced Hanabi player to play with the bot and he pointed out a few things in particular that he'd like to start using. For example, humans usually have a rule that if your teammate hints multiple cards of the same color/number, you should play the newest one. The bot uses a more complicated rule: if the card you just picked up was hinted then play that card, otherwise play the oldest hinted card. That gives you way more flexibility to hint playable cards that would otherwise be tough to get played.

I think one important general lesson is that search is really, really important. Deep RL algorithms are making huge advancements, but Deep RL alone can't reach superhuman performance in Go or poker with search. Here, too, we see that search was the key to conquering this game, and I think that will hold true in more complex real-world settings as well. Figuring out how to extend search to more complex real-world settings will be a challenge, but it's one worth pursuing.

The search algorithm shares a lot in common with our Pluribus poker AI (https://ai.facebook.com/blog/pluribus-first-ai-to-beat-pros-...), but we added "retrospective belief updates" which makes it way more scalable. We also didn't use counterfactual regret minimization (CFR) because in cooperative games you want to be as predictable as possible, whereas CFR helps make you unpredictable in a balanced way (useful in poker).

The most surprising takeaway is just how effective search was. People were viewing Hanabi as a reinforcement learning challenge, but we showed that adding even a simple search algorithm can lead to larger gains than any existing deep RL algorithm could achieve. Of course, search and RL are completely compatible, so you can combine them to get the best of both worlds, but I think a lot of researchers underestimated the value of search.

Hi! I'm one of the authors on the paper. We'd be happy to answer any questions. Ask us anything!

The humans knew the whole time which player was the bot.

The hand logs from the 5 humans + 1 AI experiment are included in the supplementary material of the Science paper.

There was real money at stake in this experiment. The pros were guaranteed $0.40 per hand just for participating, but that could increase to $1.60 per hand depending on how well they did.

To answer your question, no, I don't think human players would play at their best when not playing for actual money.

a. There was this paper a couple years ago applying CFR to single-agent settings: https://arxiv.org/abs/1710.11424

b. It really depends on the game and the situation. It can be several orders of magnitude in six-player poker. In other games, it can be even more.

Unfortunately we don't have any plans to do that currently.

We played 10,000 hands of poker in the 5 humans + 1 AI experiment. The number of hands won isn't a useful metric in poker. If you win only 10% of your hands and make $1,000 on those hands, while losing only $1 on the other 90% of hands, then you're a winning player. The bot won at a rate of 4.8 bb/100 ($4.8 per hand if the blinds are $50/$100). This is considered a large win rate by professionals.

Our goal is to make the research as accessible as possible to the AI community, so we include descriptions of the algorithms and pseudocode in the supplementary material. However, in part due to the potential negative impact this code could have on online poker, we're not releasing the code itself.

From an AI and game theory standpoint, there isn't much difference between two-team zero-sum and two-player zero-sum if the teammates are trained together. That said, the Dota 2 work is extremely impressive for a variety of other reasons.

The CFR algorithm is actually somewhat similar to Q-learning, but the connection is difficult to see because the algorithms came out of different communities, so the notation is all different.

I think the bot would make a lot of money playing against average recreational players, but it's absolutely true that if you can exploit bad players' weaknesses, then you can make more money than what the bot would earn.

We played 10,000 hands over 12 days in the 5 humans + 1 AI experiment. That's quite a long time, and there's no indication that they even began to uncover any weaknesses in that time period. So I'm fairly confident the AI is robust to exploitation, and I think that's a very important quality to have in any AI system.

Honestly, probably debugging. Training this thing is very cheap, but the variance in poker is huge (even with the best variance-reduction techniques) so it takes a very long time to tell whether one version is better than another version (or better than a human).

No, I don't have any plans to do that. This is really about advancing fundamental AI research.

It doesn't exploit its opponents' weaknesses. Its focus was on not having any weaknesses that its opponents could exploit. However, the algorithms are not guaranteed to converge to a Nash equilibrium in this setting because it's not a two-player zero-sum game (and in either case, it's not clear that playing a Nash equilibrium would provide much benefit in this setting).

The AI doesn't adapt to the opponents, and that's still an interesting challenge for AI research. That said, at the end of the day, it was making quite a bit of money playing against elite human pros. I think that suggests the cliche is, at least in part, wrong.

The pros described the bot's preflop strategy as very sensible, so I think it's unlikely there were more multiway pots with 5 AI's.

The bot bluffs, and understands that when its opponent bets it might be a bluff. I would consider that to be strategic behavior. The fact that its strategy is determined by a mathematical process doesn't change that in my opinion.

That took place after the final version of the Science paper was submitted. It would have been nice to include but it takes a while to do those experiments and we didn't feel it was worth delaying the publication process for it.