jakobnicolaus

We are entirely focused on the self-play setting in which the goal is to learn the highest performing policy for a team of agents all trained together. The Hanabi Challenge also outlines an ad-hoc setting in which you need to adjust to the diverse policies of other agents in the team on the fly.

We haven't yet analyzed the gameplay to look for examples of these well-known human Hanabi conventions. All the code and agents are open-sourced though, so feel free to take a look!

There are unique challenges around learning effective communication protocols that appear in cooperative settings, which was the focus of this work. Getting robust superhuman performance in SC2 remains an interesting challenge, though.

Absolutely, please shoot me an email. Did I mention that we link out random games our bot played in the BAD paper? Sorry for the late reply!

Sure, but in Hanabi the point is to be as informative as possible, while in poker it should be the opposite (unless you collude).

Hanabi is fully cooperative and entirely focused on communication. I think it's good to have a testbed that isolates these challenges, rather confounding them with the zero-sum (competitive) aspect of Bridge. Having said this, I do believe Bridge is a good testbed to put the different methods together.

The good news is that we have open-sourced the environment, so if you think it's easy I would love to see a simple method that solves it.

yes - this was the focus of our method: Allowing agents to interpret the actions of others, while also learning to be interpretable when observed by other agents.

Yes, but actively communicating with some of the other players through agreed conventions would probably count as collusion and be illegal in N-player poker..

Thanks for your summary of Hanabi! You can find an example of your hypothetical AI in our recent paper: https://arxiv.org/abs/1811.01458. Note that all the conventions and rules are learned though, rather than hand coded. And your 'giant lookup table', is just a feedforward neural network. We also link out 300 games in case you are curious to see how it works.

I think neural networks will be part of the solution, but they are probably not the entire answer. For an example of a method that combines Deep RL with Bayesian reasoning, you can take a look at our recent paper (https://arxiv.org/abs/1811.01458). BAD achieves best known scores in for 2 player Hanabi in self-play.

Hanabi is a multi-agent problem. Unfortunately gym doesn't natively support multi-agent action and state spaces.

A couple of years ago I built an app (PayMeMaybe) which uses this idea to settle small debts between friends: https://play.google.com/store/apps/details?id=app.paymemaybe...

Wasn't quite the break through success I suspected :)