NeuralNetPlayerA

Role: Player A implementation that uses a Keras model to map a binary game field to a communication bit-vector, optionally sampling actions and tracking the last action log-probability.

Location: Q_Sea_Battle.neural_net_player_a.NeuralNetPlayerA

Constructor

Parameter	Type	Description
game_layout	GameLayout, constraints: not specified, shape: N/A	Game layout describing field size and communication dimensions.
model_a	tf.keras.Model, constraints: callable like model_a(x, training=False) and returns logits compatible with output shape (batch, m), shape: N/A	Keras model mapping a batch of scaled field vectors (shape (batch, n2)) to per-bit logits (shape (batch, m)).
explore	bool, constraints: True enables Bernoulli sampling and False enables greedy thresholding, shape: scalar	If True, sample communication bits; if False, act greedily by thresholding probabilities.

Preconditions

• field_size, comms_size, n2, and m are not defined in this module; the constructor assumes game_layout provides whatever PlayerA requires to make decisions of length m over fields of length n2 (constraints not specified here).
• model_a must accept a np.ndarray, dtype float32, shape (1, n2) (after scaling) and return per-bit logits such that .numpy()[0] is array-like of shape (m,).

Postconditions

• self.model_a is set to model_a.
• self.explore is set to explore.
• self.last_logprob is set to None.

Errors

• Not specified.

Example

import numpy as np
import tensorflow as tf

# game_layout must be provided by the package; details not shown here.
player = NeuralNetPlayerA(game_layout=game_layout, model_a=tf.keras.Model(), explore=False)
field = np.zeros((n2,), dtype=np.int32)
msg = player.decide(field)
lp = player.get_log_prob()
player.reset()

Public Methods

decide

decide(field: np.ndarray, supp: Any | None = None) -> np.ndarray

Compute a communication vector for the given field.

Parameters

• field: np.ndarray, dtype float32 (after internal conversion), constraints: values intended in {0,1}, shape (n2,).
• supp: Any | None, constraints: currently unused, shape: N/A.

Returns

• actions: np.ndarray, dtype int {0,1}, shape (m,).

Errors

• Not specified.

Side effects

• Sets self.last_logprob to float, equal to the sum of per-bit log-probabilities for the returned action under the model logits.

Notes

• The input field is internally reshaped to np.ndarray, dtype float32, shape (1, n2) and scaled by subtracting 0.5, mapping 0 -> -0.5 and 1 -> +0.5.
• If self.explore is True, each bit is sampled independently via rnd < probs; otherwise actions are computed via probs >= 0.5.

logit_to_probs

logit_to_probs(logits: np.ndarray | float) -> np.ndarray | float

Convert logits to probabilities (wrapper around logit_to_prob).

Parameters

• logits: np.ndarray | float, constraints: not specified, shape: scalar or any shape.

Returns

• probs: np.ndarray | float, constraints: not specified, shape: same as logits.

logit_to_log_probs

logit_to_log_probs(logits: np.ndarray | float, actions: np.ndarray | float) -> np.ndarray | float

Compute per-bit log-probabilities for given actions under logits (wrapper around logit_to_logprob).

Parameters

• logits: np.ndarray | float, constraints: not specified, shape: scalar or any shape.
• actions: np.ndarray | float, constraints: intended values in {0,1}, shape: scalar or any shape broadcast-compatible with logits.

Returns

• log_probs: np.ndarray | float, constraints: not specified, shape: broadcast of logits and actions.

get_log_prob

get_log_prob() -> float

Return the log-probability of the most recent decided action.

Returns

• log_prob: float, constraints: finite-ness not specified, shape: scalar.

Errors

• Raises RuntimeError if self.last_logprob is None (i.e., decide() has not been called since the last reset()).

reset

reset() -> None

Reset internal state by clearing any stored log-probability.

Returns

• None: NoneType, shape: N/A.

Data & State

• model_a: tf.keras.Model, constraints: callable and returns logits compatible with message length m, shape: N/A.
• explore: bool, constraints: if True sample actions; if False greedily threshold at 0.5, shape: scalar.
• last_logprob: Optional[float], constraints: None before any decision or after reset(), otherwise equals the summed per-bit log-probability of the most recent action, shape: scalar.

Planned (design-spec)

• Not specified.

Deviations

• No design notes were provided; deviations not specified.

Notes for Contributors

• n2 (flattened field size) and m (number of communication bits) are implied by the docstrings and model I/O shapes but are not defined in this module; keep documentation and tests aligned with the definitions in GameLayout and PlayerA.
• The helper function _scale_field is module-private and is used to shift binary inputs from {0,1} to {-0.5,+0.5} before model inference.

Related

• Q_Sea_Battle.neural_net_player_a._scale_field (module-private helper used by decide).
• Q_Sea_Battle.logit_utilities.logit_to_prob (used by logit_to_probs).
• Q_Sea_Battle.logit_utilities.logit_to_logprob (used by logit_to_log_probs).
• Q_Sea_Battle.players_base.PlayerA (base class).

Changelog

• Not specified.