Observation Wrapper

textarena.core.ObservationWrapper(env: Env)

The ObservationWrapper is a base class for customizing the observations returned by an environment. It provides flexibility to preprocess or transform observations as needed.

Methods:

• get_observation() -> Tuple[int, Observations]: Retrieves and processes the observation for a specific player.
• observation(): Abstract method that must be implemented by subclasses to define transformation or processing of observations.

Observations returned by the environment are lists of tuples containing the sender ID and message for all events since the player’s last turn.

Example Observation Data:

# Observation at t0
[
    (-1, "You are Player 0 in the Negotiation Game [...]")
]

# Observation at t1
[
    (0, "Let's strategize to increase the total [...]")
    (-1, "Player 0 made the following offer to [...]")
    (1, "Thank you for your offer, Player 0 [...]")
    (-1, "Player 1 accepted the trade offer from Player 0.")
    (-1, "Player 1 made the following offer to [...]")
]

Available Observation Wrappers

LLMObservationWrapper

textarena.wrappers.LLMObservationWrapper(env: Env)

The LLMObservationWrapper converts environment observations into formatted strings for use with large language models (LLMs). It ensures that duplicate observations are not added to the full observations.

Example Usage:

# Example usage:
env = LLMObservationWrapper(base_env)
formatted_observation = env.observation(player_id=1, observation=[(0, "Hello"), (1, "Move forward")])

Example of Transformed Observations:

# Observation at t0
"[GAME] You are Player 0 in the Negotiation Game [...]"

# Observation at t1
"[GAME] You are Player 0 in the Negotiation Game [...] [Player 0] Let's strategize to increase the total [...] [GAME] Player 0 made the following offer to [...] [Player 1] Thank you for your offer, Player 0 [...] [GAME] Player 1 accepted the trade offer from Player 0. [GAME] Player 1 made the following offer to [...]"