TrainableAssistedPlayerB

Role: Trainable Player B wrapper that consumes Player A "previous" tensors plus local measurements to decide whether to shoot, while tracking the last action log-probability.

Location: Q_Sea_Battle.trainable_assisted_player_b.TrainableAssistedPlayerB

Constructor

Parameter	Type	Description
game_layout	Any, not specified, shape N/A	Object providing field_size and comms_size attributes used to derive \(n2 = field\_size^2\) and \(m = comms\_size\).
model_b	LinTrainableAssistedModelB, not specified, shape N/A	Trainable model called as model_b([gun_batch, comm_batch, prev_meas_batch, prev_out_batch]) to produce a shoot logit.

Preconditions

• game_layout should expose field_size and comms_size attributes convertible to int; otherwise behavior is not specified (will likely raise at runtime).
• model_b must be callable and return a Tensor compatible with shape (1, 1) when invoked from decide().

Postconditions

• self.game_layout is set to game_layout.
• self.model_b is set to model_b.
• self.parent is set to None.
• self.last_logprob_shoot is set to None.
• self.explore is set to False.

Errors

• Not specified by constructor code.

Example

from Q_Sea_Battle.trainable_assisted_player_b import TrainableAssistedPlayerB

# game_layout must provide .field_size and .comms_size; model_b must be a LinTrainableAssistedModelB
player_b = TrainableAssistedPlayerB(game_layout=layout, model_b=model_b)

Public Methods

decide(gun, comm, supp=None, explore=None)

Decide whether to shoot (0 or 1) based on gun + comm + parent.previous tensors.

Parameters

• gun: np.ndarray, dtype int, values in {0,1}, shape (n2,), where \(n2 = field\_size^2\).
• comm: np.ndarray, dtype not specified, shape (m,), where \(m = comms\_size\); values are validated for shape only (docstring notes ints in {0,1} or floats in [0,1] for DRU).
• supp: Any | None, ignored, shape N/A.
• explore: bool | None, if not None overrides self.explore, shape N/A.

Returns

• int, constraints {0,1}, shape ().

Preconditions

• self.parent is not None and self.parent.previous is not None.
• getattr(self.game_layout, "field_size") and getattr(self.game_layout, "comms_size") exist and are convertible to int.
• gun.shape == (n2,) and gun contains only 0/1.
• comm.shape == (m,).
• self.parent.previous is a tuple-like (prev_meas_list, prev_out_list) where each is a list (enforced before later normalization).
• len(prev_meas_list) >= 1 and len(prev_out_list) >= 1.

Postconditions

• Computes shoot_logit = self.model_b([gun_batch, comm_batch, prev_meas_batch, prev_out_batch]) where gun_batch has shape (1, n2) and comm_batch has shape (1, m).
• Sets self.last_logprob_shoot to the log-probability (Python float) of the returned action under the computed logits.
• Returns shoot as greedy (shoot_prob >= 0.5) if not exploring, else samples via Uniform(0,1) < shoot_prob.

Errors

• ValueError: if gun shape mismatches (n2,).
• ValueError: if gun contains values other than 0/1.
• ValueError: if comm shape mismatches (m,).
• RuntimeError: if self.parent is None or self.parent.previous is None.
• TypeError: if self.parent.previous is not (list, list) at the initial type check.
• ValueError: if either list in self.parent.previous has length < 1.

Example

import numpy as np

# Assume player_b.parent has been set and player_a has already populated parent.previous
n2 = int(player_b.game_layout.field_size) ** 2
m = int(player_b.game_layout.comms_size)

gun = np.zeros((n2,), dtype=int)
comm = np.zeros((m,), dtype=int)

shoot = player_b.decide(gun=gun, comm=comm, explore=True)
logp = player_b.get_log_prob()

get_log_prob()

Return log-probability of the last taken shoot action (as set by decide()).

Parameters

• None.

Returns

• float, constraints not specified (log-probability), shape ().

Preconditions

• self.last_logprob_shoot is not None (i.e., decide() has been called since the last reset()).

Errors

• RuntimeError: if self.last_logprob_shoot is None.

reset()

Reset internal state.

Parameters

• None.

Returns

• None, shape N/A.

Postconditions

• self.last_logprob_shoot is set to None.

Errors

• Not specified.

Data & State

• game_layout: Any, constraints not specified, shape N/A; must provide field_size and comms_size attributes used by decide().
• model_b: LinTrainableAssistedModelB, constraints not specified, shape N/A; called by decide() to produce a shoot logit tensor.
• parent: Any | None, constraints not specified, shape N/A; expected (by decide()) to provide .previous containing prior tensors from Player A.
• last_logprob_shoot: float | None, constraints not specified, shape (); updated by decide(), cleared by reset().
• explore: bool, constraints {False, True}, shape (); default False, optionally overridden per-call via decide(..., explore=...).

Planned (design-spec)

• Not specified (no design notes provided).

Deviations

• The docstring claims parent is of type TrainableAssistedPlayers and is set by TrainableAssistedPlayers.players(), but the module only types it as Any | None and does not define or enforce this contract.
• parent.previous is initially required to be (list, list) via an explicit isinstance(..., list) check, yet later code contains normalization for non-list/tuple values ("linear case: single tensor → list of length 1") that is unreachable if the earlier check fails; these two behaviors conflict.

Notes for Contributors

• Symbols used: \(n2 = field\_size^2\) and \(m = comms\_size\) are derived inside decide() from self.game_layout.
• decide() expects self.parent.previous to be populated before it is called; ensure Player A executes first in the calling sequence.
• bernoulli_log_prob_from_logits may be imported from .logit_utils or fall back to a local implementation; changing either affects last_logprob_shoot semantics.

Related

• Q_Sea_Battle.trainable_assisted_player_b.bernoulli_log_prob_from_logits (imported if available; otherwise locally defined fallback)
• Q_Sea_Battle.trainable_assisted_player_b.LinTrainableAssistedModelB (dependency)
• Q_Sea_Battle.trainable_assisted_player_b.PlayerB (base class, imported if available; otherwise fallback)

Changelog

• 0.1: Initial version (per module docstring).