fish-env

A PettingZoo AEC environment for Fish, a six-player team card game of asking, deduction and calls, and the environment of the Fish benchmark.

• Six agents, seat_0 to seat_5, in two teams of three.
• Each agent observes only what its seat may know (its own hand, the public state and the most recent event) as 648 bits, with a mask over 913 actions.
• Declaration windows after every event, decided by a seeded priority order between teams.
• Deterministic: reset(seed=...) gives the same game every time, and every game can be saved as a record and replayed exactly.
• Passes PettingZoo's own API and seed tests.
• Agent memory for training: memory_env(remembered) puts the observations each agent had just after the events it remembers (all, the latest k, or none) in its infos.

import numpy as np
from fish_env import env

game = env()
game.reset(seed=42)
rng = np.random.default_rng(0)
for agent in game.agent_iter():
    observation, reward, terminated, truncated, info = game.last()
    if terminated or truncated:
        action = None
    else:
        action = int(rng.choice(np.flatnonzero(observation["action_mask"])))
    game.step(action)

With memory, each agent's infos holds what it remembers:

import numpy as np
from fish_env import memory_env

game = memory_env(remembered=8)  # each agent remembers the last 8 events; None for all
game.reset(seed=42)
rng = np.random.default_rng(0)
for agent in game.agent_iter():
    observation, reward, terminated, truncated, info = game.last()
    past = info["memory"]  # the observations just after the events it remembers: (n, 648)
    assert past.shape[0] <= 8
    if terminated or truncated:
        action = None
    else:
        action = int(rng.choice(np.flatnonzero(observation["action_mask"])))
    game.step(action)

Install with pip install fish-env (Python 3.14 or later). The rules engine is fish-rules; baseline agents are in fish-agents. The environment guide is in the repository.

Licensed under Apache-2.0.