Hi,
I just checked some youtube video about Deepmind, and for those who doesn't know, after chess, go, and starcraft, Hanabi is the next game they are working.
This has two differents parts:
- one where the AI plays against copy of himself;
- one where the AI plays with humans.
I was wondering what both of these parts could lead.
.
About the copies of himself:
Firstly, I writed a doc recently for help people about end game which are ridiculously hard: https://docs.google.com/document/d/1jKF ... LLApk/edit .
I think it's a really good guide for understanding what is the goal of the game. This guide tell you the 4 ways to lose a game. That means, everything you do (as actions or conventions), should have in goal to reduce the chance of these factors.
Some of these factors are more likely than others (the first and third ones are unlikely), but it doesn't change that goal.
Secondly, I want to talk about hat guessing: https://github.com/Zamiell/hanabi-conve ... uessing.md .
Hat guessing is a way to play where you associate each action and each clue to numbers, and decide for an action to do for all your next players. So you get a number of numbers equal to number of players, you add them, take the modulo of it, and give a specific conventionnal clue.
Your next player try to deduce what you have choose for all the others players, do some maths, and deduce the action you wanted him to do. And he does it. All others players does the same.
While hat guessing is really in mode 'play like bots' and not too hard to play, when you want to optimise it (=make the 'actions' depending of the hands), it became really complicated.
4 players. A is you. B, C, and D are your three next players.
You want to give partial informations to all others players with your clue (and if possible valuable information), but each player see differents things. So your clue should give informations to D, to what see C about D and C, and to what see B about C and D, knowing C will know D. (Read 'Respond Order' in the last link for another explanation.)
But where it became interresting is that D can have information about what did B and C. I'm not sure this information can be exploited for A's clue but maybe.
But the choice of A should give even more informations to each players.
But at the same time, if A decide to clues three 1 in another player's hand, the AI could assume this is enough valuable information to make the clue worth it without using 'hat guessing' techniques.
...
Thirdly, I was wondering about clues like x-4-4-4 at the start of the game. That's never done in humans game because the clue doesn't achieve much. But what it achieve, is to tell that person he got three card he can easily trash if the board state doesn't seems good.
About the human part:
I was wondering how they will code it.
If at the first turn of the game, A clues red on r2 to C. How would react B?
Problem is that it's a convention, but at the same time, B need to do something to make C not plays his r2, except if we assume that red wasn't a play clue, which seems weird to me.
I hope I made you think, would really be pleased with some answers.
I just checked some youtube video about Deepmind, and for those who doesn't know, after chess, go, and starcraft, Hanabi is the next game they are working.
This has two differents parts:
- one where the AI plays against copy of himself;
- one where the AI plays with humans.
I was wondering what both of these parts could lead.
.
About the copies of himself:
Firstly, I writed a doc recently for help people about end game which are ridiculously hard: https://docs.google.com/document/d/1jKF ... LLApk/edit .
I think it's a really good guide for understanding what is the goal of the game. This guide tell you the 4 ways to lose a game. That means, everything you do (as actions or conventions), should have in goal to reduce the chance of these factors.
Some of these factors are more likely than others (the first and third ones are unlikely), but it doesn't change that goal.
Secondly, I want to talk about hat guessing: https://github.com/Zamiell/hanabi-conve ... uessing.md .
Hat guessing is a way to play where you associate each action and each clue to numbers, and decide for an action to do for all your next players. So you get a number of numbers equal to number of players, you add them, take the modulo of it, and give a specific conventionnal clue.
Your next player try to deduce what you have choose for all the others players, do some maths, and deduce the action you wanted him to do. And he does it. All others players does the same.
While hat guessing is really in mode 'play like bots' and not too hard to play, when you want to optimise it (=make the 'actions' depending of the hands), it became really complicated.
4 players. A is you. B, C, and D are your three next players.
You want to give partial informations to all others players with your clue (and if possible valuable information), but each player see differents things. So your clue should give informations to D, to what see C about D and C, and to what see B about C and D, knowing C will know D. (Read 'Respond Order' in the last link for another explanation.)
But where it became interresting is that D can have information about what did B and C. I'm not sure this information can be exploited for A's clue but maybe.
But the choice of A should give even more informations to each players.
But at the same time, if A decide to clues three 1 in another player's hand, the AI could assume this is enough valuable information to make the clue worth it without using 'hat guessing' techniques.
...
Thirdly, I was wondering about clues like x-4-4-4 at the start of the game. That's never done in humans game because the clue doesn't achieve much. But what it achieve, is to tell that person he got three card he can easily trash if the board state doesn't seems good.
About the human part:
I was wondering how they will code it.
If at the first turn of the game, A clues red on r2 to C. How would react B?
Problem is that it's a convention, but at the same time, B need to do something to make C not plays his r2, except if we assume that red wasn't a play clue, which seems weird to me.
I hope I made you think, would really be pleased with some answers.