Hanabi

This environment is part of the classic environments. Please read that page first for general information.

Creation	make("aec", "classic/hanabi-v5")
Actions	Discrete
Parallel API	Yes
Manual Control	No
Agents	agents= ['player_0', 'player_1']
Agents	2
Action Shape	Discrete(20)
Action Values	Discrete(20)
Observation Shape	(658,)
Observation Values	[0,1]

Hanabi is a 2-5 player cooperative game where players work together to form fireworks of different colors. A firework is a set of cards of the same color, ordered from 1 to 5. Cards in the game have both a color and number; each player can view the cards another player holds, but not their own. Players cannot directly communicate with each other, but must instead remove an info token from play in order to give information. Players can tell other players which of the cards in their hand is a specific color, or a specific number. There are initially 8 info tokens, but players can discard cards in their hand to return an info token into play. Players can also play a card from their hand: the card must either begin a new firework or be appended in order to an existing firework. However, 2 fireworks cannot have the same color, and a single firework cannot repeat numbers. If the played card does not satisfy these conditions, a life token is placed. The game ends when either 3 life tokens have been placed, all 5 fireworks have been completed, or all cards have been drawn from the deck. Points are awarded based on the largest card value in each created firework.

Environment arguments

Hanabi takes in a number of arguments defining the size and complexity of the game. Default is a full 2 player hanabi game.

from pettingzoo import make

make("aec", "classic/hanabi-v5", colors=5, ranks=5, players=2, hand_size=5,
max_information_tokens=8, max_life_tokens=3, observation_type="minimal")

colors: Number of colors the cards can take (affects size of deck)

ranks: Number of ranks the cards can take (affects size of deck)

hand_size: Size of player’s hands. Standard game is (4 if players >= 4 else 5)

max_information_tokens: Maximum number of information tokens (more tokens makes the game easier by allowing more information to be revealed)

max_life_tokens: Maximum number of life tokens (more tokens makes the game easier by allowing more information to be revealed)

observation_type: “minimal”: Minimal observation (what a human sees). “card_knowledge”: includes per-card knowledge of past hints, as well as simple inferred knowledge of the form “this card is not red, because it was not revealed as red in a past”. “seer” shows all cards, including the player’s own cards, regardless of what hints have been given.

Observation Space

The observation is a dictionary which contains an 'observation' element which is the usual RL observation described below, and an 'action_mask' which holds the legal moves, described in the Legal Actions Mask section.

The main observation space of an agent is a 658 sized vector (for the default 2-player full game) representing the life and info tokens left, the currently constructed fireworks, the hands of all other agents, the current deck size and the discarded cards. Since v5 the observation is produced by OpenSpiel’s canonical Hanabi encoder (via Shimmy), so its layout — and the index ranges in the table below — depend on the chosen configuration (colors, ranks, players, hand_size, max_information_tokens, max_life_tokens). The table below is for the default configuration (colors=5, ranks=5, players=2, hand_size=5, max_information_tokens=8, max_life_tokens=3).

Each card is encoded with a colors*ranks bit one-hot vector (25 bits by default), where the index of the set bit is color*ranks + rank. The remaining deck size is represented with a unary encoding whose length is max_deck_size - players*hand_size (40 bits by default, since the players*hand_size cards dealt at the start can never be in the deck — note this is not the full deck size of 50). The state of each colored firework is represented with a one-hot encoding of the highest rank played (all-zero if no card of that color has been played yet). The information tokens remaining and the life tokens remaining are each represented with a unary encoding. The discard pile is represented with a thermometer encoding of the ranks of each discarded card, that is the least significant bit being set to 1 indicates the lowest rank card of that color has been discarded.

As players reveal info about their cards, the information revealed per card is also observed (the “card knowledge” section). Each card uses colors*ranks + colors + ranks bits (35 by default). The first colors*ranks bits represent whether or not that specific card could still be a specific color/rank given the hints received. The next colors bits store whether the color of that card was explicitly revealed, so if the card was revealed to be red, then these bits would be 10000. Finally the last ranks bits are the revealed rank of the card. So if the card was revealed to be of rank 1, then these bits would be 10000. These bits are tracked and observed for every card in every player’s hand (this player first, then the others).

Index	Description	Values
0-124	Other player’s hand: 5 cards × 25-bit one-hot	[0, 1]
125-126	Per-player flag set when that player’s hand is missing a card	[0, 1]
127-166	Unary Encoding of Remaining Deck Size (max_deck − players×hand)	[0, 1]
167-171	Vector of Red Firework	[0, 1]
172-176	Vector of Yellow Firework	[0, 1]
177-181	Vector of Green Firework	[0, 1]
182-186	Vector of White Firework	[0, 1]
187-191	Vector of Blue Firework	[0, 1]
192-199	Unary Encoding of Remaining Info Tokens	[0, 1]
200-202	Unary Encoding of Remaining Life Tokens	[0, 1]
203-252	Thermometer Encoding of Discard Pile	[0, 1]
253-254	Last action: Acting Player (relative offset, one-hot)	[0, 1]
255-258	Last action: Move Type (play / discard / reveal color / rank)	[0, 1]
259-260	Last action: Target Player of a reveal (relative offset)	[0, 1]
261-265	Last action: Color Revealed	[0, 1]
266-270	Last action: Rank Revealed	[0, 1]
271-275	Last action: Which Cards in the Hand were Revealed	[0, 1]
276-280	Last action: Position of the Card that was played or discarded	[0, 1]
281-305	Last action: Vector Representing the Card that was played/discarded	[0, 1]
306-306	Last action: Whether the played card was added to a firework	[0, 1]
307-307	Last action: Whether an info token was added	[0, 1]
308-342	Card Knowledge of This Player’s 0th Card	[0, 1]
343-377	Card Knowledge of This Player’s 1st Card	[0, 1]
378-412	Card Knowledge of This Player’s 2nd Card	[0, 1]
413-447	Card Knowledge of This Player’s 3rd Card	[0, 1]
448-482	Card Knowledge of This Player’s 4th Card	[0, 1]
483-517	Card Knowledge of Other Player’s 0th Card	[0, 1]
518-552	Card Knowledge of Other Player’s 1st Card	[0, 1]
553-587	Card Knowledge of Other Player’s 2nd Card	[0, 1]
588-622	Card Knowledge of Other Player’s 3rd Card	[0, 1]
623-657	Card Knowledge of Other Player’s 4th Card	[0, 1]

Legal Actions Mask

The legal moves available to the current agent are found in the action_mask element of the dictionary observation. The action_mask is a binary vector where each index of the vector represents whether the action is legal or not. The action_mask will be all zeros for any agent except the one whose turn it is. Taking an illegal move ends the game with a reward of -1 for the illegally moving agent and a reward of 0 for all other agents.

Action Space

The action space is a scalar value, which ranges from 0 to the max number of actions. The values represent all possible actions a player can make, legal or not. Each possible move in the environment is mapped to a UUID, which ranges from 0 to the max number of moves. By default the max number of moves is 20. The first range of actions are to discard a card in the agent’s hand. If there are k cards in the player’s hand, then the first k action values are to discard one of those cards. The next k actions would be to play one of the cards in the player’s hand. Finally, the remaining actions are to reveal a color or rank in another players hand. The first set of reveal actions would be revealing all colors or values of cards for the next player in order, and this repeats for all the other players in the environment.

Action ID	Action
0	Discard Card at position 0
1	Discard Card at position 1
2	Discard Card at position 2
3	Discard Card at position 3
4	Discard Card at position 4
5	Play Card at position 0
6	Play Card at position 1
7	Play Card at position 2
8	Play Card at position 3
9	Play Card at position 4
10	Reveal Red Cards for Player 1
11	Reveal Yellow Cards for Player 1
12	Reveal Green Cards for Player 1
13	Reveal White Cards for Player 1
14	Reveal Blue Cards for Player 1
15	Reveal Rank 1 Cards for Player 1
16	Reveal Rank 2 Cards for Player 1
17	Reveal Rank 3 Cards for Player 1
18	Reveal Rank 4 Cards for Player 1
19	Reveal Rank 5 Cards for Player 1

Rewards

The reward of each step is calculated as the change in game score from the last step. The game score is calculated as the sum of values in each constructed firework. If the game is lost, the score is set to zero, so the final reward will be the negation of all reward received so far.

For example, if fireworks were created as follows:

Blue 1, Blue 2, Red 1, Green 1, Green 2, Green 3

At the end of the game, the total score would be 2 + 1 + 3 = 6

If an illegal action is taken, the game terminates and the one player that took the illegal action loses. Like an ordinary loss, their final reward will be the negation of all reward received so far. The reward of the other players will not be affected by the illegal action.

Version History

• v5: Switched environment to depend on OpenSpiel (using Shimmy) for future compatibility (1.23.0)

• v4: Fixed bug in arbitrary calls to observe() (1.8.0)

• v3: Legal action mask in observation replaced illegal move list in infos (1.5.0)

• v2: Fixed default parameters (1.4.2)

• v1: Bumped version of all environments due to adoption of new agent iteration scheme where all agents are iterated over after they are done (1.4.0)

• v0: Initial versions release (1.0.0)

Usage

AEC

from pettingzoo import make

env = make("aec", "classic/hanabi-v5", render_mode="human")
env.reset(seed=42)

for agent in env.agent_iter():
    observation, reward, termination, truncation, info = env.last()

    if termination or truncation:
        action = None
    else:
        mask = observation["action_mask"]
        # this is where you would insert your policy
        action = env.action_space(agent).sample(mask)

    env.step(action)
env.close()

API

class pettingzoo.classic.hanabi.hanabi.env(**kwargs)

class pettingzoo.classic.hanabi.hanabi.raw_env(colors: int = 5, ranks: int = 5, players: int = 2, hand_size: int = 5, max_information_tokens: int = 8, max_life_tokens: int = 3, observation_type: str = 'card_knowledge', random_start_player: bool = False, render_mode: str | None = None)

Initializes the raw_env class.

Parameter descriptions :

• colors: int, Number of colors in [2,5].

• ranks: int, Number of ranks in [2,5].

• players: int, Number of players in [2,5].

• hand_size: int, Hand size in [2,5].

• max_information_tokens: int, Number of information tokens (>=0).

• max_life_tokens: int, Number of life tokens (>=1).

• observation_type: str.

  “minimal””: Minimal observation. “card_knowledge”: First-order common knowledge observation. “seer”: Full information of all cards.

• random_start_player: bool, Random start player.

Common game configurations:

Hanabi-Full (default){

“colors”: 5, “ranks”: 5, “players”: 2, “max_information_tokens”: 8, “max_life_tokens”: 3, “hand_size”: (4 if players >= 4 else 5) “observation_type”: “card_knowledge”, “hand_size”: 2, }

Hanabi-Small{

“colors”: 2, “ranks”: 5, “players”: 2, “max_information_tokens”: 3, “hand_size”: 2, “max_life_tokens”: 1, “observation_type”: “card_knowledge”, }

Hanabi-Very-Small{

“colors”: 1, “ranks”: 5, “players”: 2, “max_information_tokens”: 3, “hand_size”: 2, “max_life_tokens”: 1, “observation_type”: “card_knowledge”, }

action_space(agent)

Takes in agent and returns the action space for that agent.

MUST return the same value for the same agent name

Default implementation is to return the action_spaces dict

close()

Closes any resources that should be released.

Closes the rendering window, subprocesses, network connections, or any other resources that should be released.

observation_space(agent)

Takes in agent and returns the observation space for that agent.

MUST return the same value for the same agent name

Default implementation is to return the observation_spaces dict

observe(agent_name: str)

Returns the observation an agent currently can make.

last() calls this function.

render()

Prints player’s data.

Supports console print only.

reset(seed=None, options=None)

Resets the environment for a new game and returns observations of current player as List of ints.

Returns:

• observation – Optional list of integers of length self.observation_vector_dim, describing observations of

• current agent (agent_selection)

step(action: int, observe: bool = True, as_vector: bool = True) → ndarray | list[list[dict]] | None

Advances the environment by one step. Action must be within self.legal_moves, otherwise throws error.

Returns:

• observation – Optional List of new observations of agent at turn after the action step is performed.

• By default, a list of integers, describing the logic state of the game from the view of the agent.

• Can be a returned as a descriptive dictionary, if as_vector=False.