Hanabi

Robert B. Kaspar, [email protected]

All contributors welcome. Please maintain compatibility with both Python 2 and Python 3.

Usage

usage: ./hanabi_wrapper.py p1 p2 [p3 ...] [options]
  pi (AI for player i): idiot, cheater, basic, brainbow, newest, encoder, gencoder, hat, or human
  -t game_type: rainbow [default], purple (6 suits), vanilla (5 suits), or black (6 suits; one suit has only 1 copy of each rank)
  -n n_rounds: positive int [default: 1]
  -v verbosity: verbose [default], scores, silent, or log
  -l loss_score (points to award after 3 guesses): zero [default] or full
  -s seed: if given, use this number as the seed for the RNG
  -p: if given, try to check whether a bot cheats, and gives an error if it does.
  -o output: if given, output a JSON file of the game in log.json

There is no max number of players. With >5, hand size is still 4 cards.

Example usage

$ ./hanabi_wrapper.py newest newest newest newest

or

$ ./hanabi_wrapper.py cheater cheater -t purple -n 1000 -v silent

Example output

ROUND 0:
[HANDS] Newest1: 1g 1? 2g 4?
        Newest2: 3g 3b 1b 4b
        Newest3: 3y 2r 3w 4r
        Newest4: 3r 4r 5y 1r
[PLAYS] Newest1 [1g 1? 2g 4?] hints 1 to Newest2
        Newest2 [3g 3b 1b 4b] plays 1b and draws 1g
        Newest3 [3y 2r 3w 4r] hints 1 to Newest4
        Newest4 [3r 4r 5y 1r] plays 1r and draws 3r
        ...
        Newest1 [4y 4b 2w 1?] plays 4b and draws 1g
        Newest2 [1r 2y 4y 1?] discards 1r and draws 1w
        Newest3 [5? 1b 1y 3y] hints 5 to Newest4
        Newest4 [1b 5b 4g 2?] plays 5b
        Newest1 [4y 2w 1? 1g] hints 3 to Newest3
        Newest2 [2y 4y 1? 1w] discards 2y
Score: 25

or

AVERAGE SCORE: 29.81 +/- 0.02 (1 std. err.)
PERFECT GAMES: 88.6%

Available players

• Cheating Idiot (idiot) by RK
  Peeks at own hand to know when to play, discards randomly, never hints
• Cheater (cheater) by Floris van Doorn
  Peeks at own hand to play, discard, and hint; high win rate
• Most Basic (basic) by Ben Zax
  Plays when certain, discards randomly, hints inefficiently, no rainbows
• Basic Rainbow (brainbow) by Greg Hutchings
  Like basic but checks direct and indirect info to handle rainbows
• Newest Card (newest) by BZ
  Plays newest hinted card (and hints accordingly), discards oldest card
• Encoding (encoder) & General Encoding (gencoder) by Taylor Robie
  Experimental, hints counter-intuitively, Python 2 only (todo: 3!)
• Hat Player (hat) by FvD
  Uses "hat guessing" techniques to convey information to all other players with a single hint. Needs at least 4 players. Has about 94% win rate. See documentation.
• Human (human) by GH
  Allows you to play alongside the AIs (works best on -v silent or log)

Optimized Winrates

There are two optimized players, with high winrates.

The hat player gets the following winrates on the different variants (it cannot play with 2 or 3 players):

Players	vanilla	6 suits	rainbow	black
4	94.1	93.9	93.7	59.6
5	90.8	95.1	95.1	56.5

The cheating player gets the following winrates on the different variants. The cheating player can look at their own hand and play cards without needing clues. It does not look in the deck.

Players	5 suits	6 suits	black
2	94.8	90.2	67.3
3	98.4	98.4	79.8
4	98.1	98.3	76.7
5	96.8	97.7	72.6

The average score tables can be found in the source files themselves.

How to write your own AI player

Use an existing player as a guide. CheatingIdiot is especially simple.

Just make a player class with a play method whose only argument is a Round instance. (Round stores all of the game information for a single round.)

play must return a two-tuple. The 1st entry tells the framework what kind of action your player is taking: 'hint', 'play', or 'discard'. The 2nd entry specifies the action's target. For a play or discard, that's just which card to use. For a hint, it's another two-tuple: the target player (int between 0 and nPlayers - 1) and the info to give (a one-char str representing a color or number). See Round.get_play() in hanabi_classes.py for more info.

You'll want to use the information available to you from other players' hands, the tableau, the discard pile, and how many hints are left to inform your AI's choices. This info is available in the Round object; see especially Round.__init__(). Note that player hands are stored as sub-objects of Round. For example, in Round instance r a list of player i's cards is available as r.h[i].cards. (Don't look at your own cards unless you're despicable like CheatingIdiot! ... You make me sick.)

Also add your class to the README.

Installation if needed

If you need to install dependencies to make the project run:

pip install -r requirements.txt

It is however recommended you do run this inside of a virtualenv, like below:

virtualenv venv
./venv/bin/activate
pip install -r requirements.txt
./hanabi_wrapper.py newest newest newest newest

And then when you're done - you can exit the virtualenv:

deactivate