Visualization

2 min read

Each plot answers a single question. They share the resource’s six semantic colours, so a figure in a notebook and a diagram on the website agree about what a colour means.

Import from cooperative_marl_labs.visualization.

Functions

plot_protocol_heatmap

from cooperative_marl_labs.visualization import plot_protocol_heatmap

plot_protocol_heatmap(
    matrix: np.ndarray,
    target_names: list[str] | None = None,
    title: str = 'Learned protocol  P(m | target)',
    ax=None,
)

One row per target, one column per symbol.

A converged protocol has exactly one bright cell per row. A row spread evenly across several columns means the speaker never committed to a convention for that target.

plot_crossplay_matrix

from cooperative_marl_labs.visualization import plot_crossplay_matrix

plot_crossplay_matrix(
    df: pd.DataFrame,
    title: str = 'Cross-play',
    ax=None,
    vmax: float | None = None,
)

Every ego against every partner, with the familiar pairings outlined.

The diagonal is what a standard evaluation reports. The off-diagonal is the number that says whether anything generalized, so read the two together or the figure says nothing.

plot_partner_estimate

from cooperative_marl_labs.visualization import plot_partner_estimate

plot_partner_estimate(
    estimates,
    truth=None,
    switch_at: int | None = None,
    title: str = 'What the ego believes about its partner',
    ax=None,
)

An estimator’s belief over time, against the truth it is chasing.

Parameters

estimates: Estimated P(FETCH) after each step. truth: The partner’s actual P(FETCH), a scalar or one value per step. switch_at: Step at which the partner was replaced, marked with a vertical rule. A windowed estimator crosses to the new value; one that averages the whole history does not.

render_wireless_network

from cooperative_marl_labs.visualization import render_wireless_network

render_wireless_network(
    env,
    actions=None,
    title: str = 'Channel allocation',
    ax=None,
    show_users: bool = True,
)

Draw the current allocation.

Parameters

env: A WirelessResourceAllocationEnv. Its positions, demands and coupling matrix are what get drawn. actions: Channel per access point. Defaults to the environment’s last step.

plot_wireless_comparison

from cooperative_marl_labs.visualization import plot_wireless_comparison

plot_wireless_comparison(
    results: dict[str, dict[str, float]],
    metric: str = 'team_reward',
    ceiling: float | None = None,
    title: str | None = None,
    ax=None,
)

One bar per system, in the order given.

Parameters

results: {system name: metrics dict}, as returned by evaluate_agents. ceiling: Best achievable value, drawn as a rule. A bar chart without it invites the reader to treat the tallest bar as good rather than as best so far.

channel_colour

from cooperative_marl_labs.visualization import channel_colour

channel_colour(channel: int) -> str

The colour for one channel, wrapping if there are more than five.

The same three colours the lab illustration uses, so a plot and the picture on the website agree about which channel is which.

Constants

Name	Value
`CHANNEL_COLOURS`	`('#5448c8', '#2191fb', '#5da271', '#fea82f', '#fc5130')`