Match Environment¶

gwent.rl.pettingzoo exposes the match-only environment.

from gwent.rl.pettingzoo import env, raw_env

env(...) wraps the raw environment with PettingZoo's AssertOutOfBoundsWrapper and OrderEnforcingWrapper. raw_env(...) returns GwentPettingZooEnv directly.

Constructor¶

raw_env(
    *,
    decks=None,
    seed=0,
    first_player=0,
    strict_effects=True,
    catalog_version=STANDALONE_VERSION,
    render_mode=None,
)

Option	Meaning
`decks`	Optional tuple of two `DeckList` objects. Defaults to the shared demo deck for mirror testing.
`seed`	Seed used to create the core deterministic RNG.
`first_player`	`0` or `1`; invalid values raise `ValueError`.
`strict_effects`	Passed into core match setup and deck validation.
`catalog_version`	Card catalog version used for categorical card ids.
`render_mode`	`None` or `"ansi"`.

reset(options=...) can override decks, first_player, and strict_effects for that reset.

Agents And Spaces¶

The possible agents are always:

("player_0", "player_1")

Each agent has:

Discrete(349) action space.
Dict({"observation": MultiDiscrete(...), "action_mask": MultiBinary(349)}) observation space.

The current generated reference reports the exact observation schema and action layout:

Rewards¶

Rewards are sparse and terminal:

Result	`player_0`	`player_1`
`player_0` wins	`+1.0`	`-1.0`
`player_1` wins	`-1.0`	`+1.0`
Draw	`0.0`	`0.0`
Non-terminal step	`0.0`	`0.0`

At game over both agents terminate, neither truncates, and PettingZoo dead-step handling removes agents from the active list.

Infos¶

infos[agent] contains:

legal_action_ids: the sorted legal discrete action ids for the active live agent, otherwise an empty list.
events: serializable event dictionaries from the last core engine step.

Rendering¶

render_mode="ansi" returns a compact text snapshot with phase, round, active agent, scores, player counts, row contents, and pending-choice count. render_mode=None disables rendering.