AI and 'hat guessing'

[Romain672]

Group for those who wanna try it: https://boardgamearena.com/group?id=955 ... ublicinfos

Hi,
In hanabi, the hat guessing consist to attribute a value to each clue, and to give a piece of information to all others players.
Per example, if the table is:
value | action given | hint to give
0 | give clue | white to last player
1 | play slot 1 | 1 to next player
2 | play slot 2 | 2 to next player
3 | play slot 3 | 3 to next player
4 | play slot 4 | 4 to next player
5 | play slot 5 | 5 to next player
6 | discard slot 1 | red to next player
7 | discard slot 2 | yellow to next player
8 | discard slot 3 | green to next player
9 | discard slot 4 | blue to next player
10 | discard slot 5 | white to next player
11 | - | 1 to second next player
and so on...

And the starting hands are:
b1-g4-w4-g5-b5 for your next player
g4-y3-b2-y3-w3 for your second next player

You will associate an action to do for your previous player (known by all others players), here it would be discard slot 2* because y3 is duplicate. Discard slot 2 has a value of 7.
And then associate an action to do for your second previous player, here it would be play slot 1. Play slot 1 has a value of 1.
You add all values: 7+1=8.
Then you give the clue associated to that value: green to your next player.

THEN:
Your next player check the value for green to your next player: 8.
Your next player check the action decided for your previous player, which is discard slot 2. Value of 7.
He does 8-7=1. He does action '1' so play slot 1.

THEN:
Your previous player check the value for green to your next player: 8.
He check the action done by the other player: play slot 1. Value of 1.
He does 8-1=7. He does action '7' so discard slot 2.

* Why slot 2 and not 5? Because it's leftmost. It must be decided conventionnally before the game for all players to be agree on what is going on.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~

I was wondering if this kind of reasonning can be applicated to Tranquility.
I believe the most simple thing to work around is with a manual ship: you will have the choice between 5 differents places.
You could do something like that: you attribute the value of 1 on the lowest place the ship can go, 2 just above, and so on. So you can have value from 1 to 5.
Let's assume you always communicate your lowest card during the first turn, and don't play it (if you draw a lower card at the end of your turn, you ignore it).
The first move you do for the ship will be associated with the (value of the place-1)*5.
The second move you do for the ship will be associated with the value of the place.
You add both of them and:
- if it's a number between 1 and 24, it's the lowest card you got during the first turn
- if it's 25, that means all your cards during your first turn got a value of 25 or more

This is the base which is the easiest thing I can do.
It look like telling the lowest number isn't good, telling a number between 28 and 51 is best but that's more difficult (or 30 to 53 for simplicity).
And I'm not sure what you do during turns 3/4. You want to tell another number. But which one.
And how powerful that info is.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Here is another possibility:
You got value from 1 to 5 with the ship.
If you choose a number from 1 to 4, that means you got the card which is 1/2/3/4 higher than the card you just played. If you choosed 5, that means you didn't.

That's maybe way easier.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~

You could argue: you will often put the ship in the highest line, which is bad, and I agree.
One way to work around that is to cycle the 1/2/3/4/5 spots. That means during the first turn you got the normal value.
But during the second you got the value 2/3/4/5/1. Then 3/4/5/1/2. And so on.

Thoughs? Ideas?

[Romain672]

I tried two games yesterday.
The first one we associate value of +1/+2/+3/+4/none to the ship, and there was lots of hints which didn't do much.
The second one we associate value of +1/+2/none/-2/-1 to the ship, and we did nearly no hints during the whole (because the +1/+2 is fully symetrical of the -1/-2 early on).

From that, it seems regrouping +1 or +2 in a single value is great: it tell the other player he can play a +2/+3 safely next to it.
And going in negatives doesn't seems great.


The goal seem to only give valuable information.
Here is a board which regroup all the information you want to send: https://www.zupimages.net/up/21/19/zpis.jpg
My assumptions were:
- you always wanna play a +4/+5 if you can (2steps after)(which means if with the info send you don't change what you planned to do (=playing a +4/+5) this info will be useless, so grey)
- you don't want to play a +3 unless you have nothing better to do (2steps after)
- you don't want to play a +8 unless you have nothing better to do (4steps after)
- you want to play a +9/+10 if you can (4steps after)
- you could want to play a +7 4steps after if you know the other player got a +1/+3/+4/+6

Per example, if you tell the other player you got a '+2' (so line +2) and:
- the other player got a +1, it's pink (meh)
- the other player got a +3, it's great (green)
- and all others are useless

What count the most are the line which got some green. So lines +1/+2/+7 in priority.
You can regroup +1 and +2 because even if the other player doesn't know which one it is, it doesn't change much things.
It seems after the +7 is really good because it open a +6 play. But that's pretty much it.
I dunno what to said after, there is lots of lines which can be situationnally useful (I'm unsure the orange is better than the blue).


So here is the thing I want to try next:
Value 1): +1 or +2
Value 2): +4 or +5
Value 3): +6 or +7
Value 4): +11 or +12
Value 5): none of the above

But feel free to react to it or tell me if you disagree with (mainly) the assumptions.

If you want another layer of complexity, you can sometimes ask a question. Per example if someone tell he got the +2, he would really want in the next turn to let you said yes or no in his answer. But that complexify a lot the conventions.

[ufm]

Hat guessing assumes 'you can look all hands but your own', and hinter clueing the other players their hands by the play.
However Tranquility is a standard tile placement game.
For me this looks like an ordinary convention, not hat guessing.

[Romain672]

Just created a group for those who wanna try it: https://boardgamearena.com/group?id=955 ... ublicinfos

But yeah I agree hat guessing isn't a good name