Partner Policies

1 min read

Each partner takes the FETCH role with its own probability, so no single observation identifies one. Telling them apart needs a history, which is what makes a partner model worth building.

Import from cooperative_marl_labs.policies.

About this module

Scripted partner policies for the Adapt practical.

Classes

PartnerPolicy

from cooperative_marl_labs.policies import PartnerPolicy

PartnerPolicy(seed: int | None = None) -> None

Base class for a scripted partner.

Owns a seeded generator, so a partner’s behaviour is reproducible and no partner touches global NumPy random state.

Methods

• act(observation: Any = None, history: Any = None) -> int
• reset() -> None
  Called at every env.reset. Stateless partners need nothing.

FetchFirstPartner

from cooperative_marl_labs.policies import FetchFirstPartner

FetchFirstPartner(seed: int | None = None) -> None

Almost always fetches. The ego should cook.

Methods

• act(observation: Any = None, history: Any = None) -> int
• reset() -> None
  Called at every env.reset. Stateless partners need nothing.

CookFirstPartner

from cooperative_marl_labs.policies import CookFirstPartner

CookFirstPartner(seed: int | None = None) -> None

Almost always cooks. The ego should fetch.

Methods

• act(observation: Any = None, history: Any = None) -> int
• reset() -> None
  Called at every env.reset. Stateless partners need nothing.

BalancedPartner

from cooperative_marl_labs.policies import BalancedPartner

BalancedPartner(seed: int | None = None) -> None

Takes each role half the time, so nothing the ego does beats 0.5.

Methods

• act(observation: Any = None, history: Any = None) -> int
• reset() -> None
  Called at every env.reset. Stateless partners need nothing.

ReactivePartner

from cooperative_marl_labs.policies import ReactivePartner

ReactivePartner(seed: int | None = None) -> None

Takes whichever role the ego did not take last step.

Unlike the others this one is not stochastic, so an ego that settles into a fixed role is complemented perfectly. It is in the training population to stop a learner assuming every partner is a coin flip.

Methods

• act(observation: Any = None, history: Any = None) -> int
• reset() -> None
  Called at every env.reset. Stateless partners need nothing.

HeldOutPartner

from cooperative_marl_labs.policies import HeldOutPartner

HeldOutPartner(seed: int | None = None) -> None

Cook-leaning, and never used during training.

Methods

• act(observation: Any = None, history: Any = None) -> int
• reset() -> None
  Called at every env.reset. Stateless partners need nothing.

Functions

all_partners

from cooperative_marl_labs.policies import all_partners

all_partners(seed: int | None = 0) -> dict[str, PartnerPolicy]

Every partner, keyed by name, each with its own offset seed.

training_partners

from cooperative_marl_labs.policies import training_partners

training_partners(seed: int | None = 0) -> dict[str, PartnerPolicy]

Just the partners the ego may train against.

held_out_partners

from cooperative_marl_labs.policies import held_out_partners

held_out_partners(seed: int | None = 0) -> dict[str, PartnerPolicy]

Just the evaluation-only partners.

Constants

Name	Value
TRAINING_POPULATION	('fetch-first', 'cook-first', 'balanced', 'reactive')
HELD_OUT	('held-out',)