LinTrainableAssistedModelB

Role: Player B linear trainable-assisted baseline model that consumes Player A’s previous measurement and PR-assisted outcome tensors to produce a shoot logit.

Location: Q_Sea_Battle.lin_trainable_assisted_model_b.LinTrainableAssistedModelB

Derived constraints

Define field_size as the side length of the square field, comms_size as the size of the communication vector, and n2 = field_size * field_size as the flattened field length.

Constructor

Parameter	Type	Description
field_size	int, constraint: \(field\_size \ge 1\)	Side length of the square field; used to compute n2 = field_size * field_size.
comms_size	int, constraint: \(comms\_size \ge 1\)	Length of the communication vector consumed by the combine layer.
sr_mode	str, constraint: unspecified allowed values, default "expected"	Mode passed to PRAssistedLayer as mode. If the resulting self.pr_assisted.mode == "sample", measurements are thresholded to binary via meas_probs_b >= 0.5.
seed	int | None, constraint: unspecified, default 0	Seed passed to PRAssistedLayer.
p_high	float, constraint: unspecified, default 0.9	p_high passed to PRAssistedLayer.
resource_index	int, constraint: unspecified, default 0	resource_index passed to PRAssistedLayer.
hidden_units_meas	Sequence[int], constraint: unspecified, default (64,)	Hidden layer configuration passed to LinMeasurementLayerB(hidden_units=...).
hidden_units_combine	int | Sequence[int], constraint: unspecified, default (64, 64)	Hidden layer configuration passed to LinCombineLayerB(hidden_units=...).
name	str | None, constraint: unspecified, default None	Keras model name; if None, uses "LinTrainableAssistedModelB".
**kwargs	Any, constraint: unspecified	Forwarded to tf.keras.Model.__init__.

Preconditions: field_size and comms_size must be convertible to int.

Postconditions: Sets self.field_size: int, self.comms_size: int, self.n2: int; constructs self.measurement: LinMeasurementLayerB, self.pr_assisted: PRAssistedLayer, self.combine: LinCombineLayerB; also defines backward-compatible aliases measure_layer, pr_layer, resource_layer, sr_layer, combine_layer referencing those layers.

Errors: Not specified (aside from potential errors raised by sublayer constructors or invalid type conversions).

Example

import tensorflow as tf
from Q_Sea_Battle.lin_trainable_assisted_model_b import LinTrainableAssistedModelB

model = LinTrainableAssistedModelB(field_size=5, comms_size=3, sr_mode="expected")

B = 2
n2 = 25
gun = tf.zeros((B, n2), dtype=tf.float32)
comm = tf.zeros((B, 3), dtype=tf.float32)
prev_meas_list = [tf.zeros((B, n2), dtype=tf.float32)]
prev_out_list = [tf.zeros((B, n2), dtype=tf.float32)]

shoot_logit = model([gun, comm, prev_meas_list, prev_out_list], training=False)

Public Methods

_ensure_batched

Static method: LinTrainableAssistedModelB._ensure_batched(x) -> (x_batched, already_batched)

Parameters:

• x: tf.Tensor, dtype convertible to float32, shape (n2,) or (B, n2) or any tensor where rank is used to decide batching; rank-1 inputs are expanded on axis 0.

Returns:

• x_batched: tf.Tensor, dtype float32, shape (1, N) if input rank is 1, else same shape as input (rank not modified); N is the input’s last-dimension size when rank-1.
• already_batched: bool, constraint: False if input rank is 1 else True.

Errors: Not specified.

compute_with_internal

Method: compute_with_internal(gun_batch, comm_batch, prev_meas_list, prev_out_list, training=False) -> (shoot_logit, meas_list, out_list)

Parameters:

• gun_batch: tf.Tensor, dtype float32, shape (B, n2) (rank-1 shape (n2,) is accepted and converted to (1, n2)); last dimension must equal n2.
• comm_batch: tf.Tensor, dtype float32, shape (B, comms_size) (rank-1 shape (comms_size,) is accepted and converted to (1, comms_size)); last dimension must equal comms_size.
• prev_meas_list: List[tf.Tensor] (or tf.Tensor accepted and normalized to a list), dtype float32, shape list length \(\ge 1\) where element 0 has shape (B, n2).
• prev_out_list: List[tf.Tensor] (or tf.Tensor accepted and normalized to a list), dtype float32, shape list length \(\ge 1\) where element 0 has shape (B, n2).
• training: bool, constraint: unspecified, default False; forwarded to self.measurement(...) and self.combine(...).

Returns:

• shoot_logit: tf.Tensor, dtype float32, shape (B, 1) (if combine returns rank-1 (B,), it is expanded to (B, 1)).
• meas_list: List[tf.Tensor], constraint: list length 1, containing meas_for_resource_b: tf.Tensor, dtype float32, shape (B, n2); if self.pr_assisted.mode == "sample", this tensor is binary-valued in {0.0, 1.0} via thresholding at 0.5, otherwise it equals meas_probs_b cast to float32.
• out_list: List[tf.Tensor], constraint: list length 1, containing outcomes_b: tf.Tensor, dtype not specified by this module, shape (B, n2) as produced by self.pr_assisted(...).

Errors:

• Raises ValueError if comm_batch is not rank-2 after batching normalization, or if its last dimension is not comms_size.
• Raises ValueError if gun_batch is not rank-2 after batching normalization, or if its last dimension is not n2.
• Raises ValueError if prev_meas_list or prev_out_list is empty after normalization to list/tuple.
• Raises ValueError if prev_meas or prev_out is not rank-2, or if their last dimensions are not n2.
• Raises ValueError if meas_for_resource_b is not rank-2 or its last dimension is not n2.

PR-assisted invocation contract

This method calls self.pr_assisted with a dict containing keys "current_measurement", "previous_measurement", "previous_outcome", and "first_measurement" where "first_measurement" is a zeros tensor of shape (B, 1) and dtype float32.

call

Method: call(inputs, training=False, **kwargs) -> shoot_logit

Parameters:

• inputs: list, constraint: must be a list/tuple of length 4 in the order [gun_batch, comm_batch, prev_meas_list, prev_out_list]; element constraints are as in compute_with_internal.
• training: bool, constraint: unspecified, default False; forwarded to compute_with_internal.
• **kwargs: Any, constraint: accepted but not used by this implementation.

Returns:

• shoot_logit: tf.Tensor, dtype float32, shape (B, 1).

Errors:

• Raises ValueError if inputs is not a list/tuple of length 4.
• Propagates ValueError from compute_with_internal for shape/rank violations.

Data & State

Instance attributes created by the constructor:

• field_size: int, constraint: \(field\_size \ge 1\), scalar.
• comms_size: int, constraint: \(comms\_size \ge 1\), scalar.
• n2: int, constraint: \(n2 = field\_size * field\_size\), scalar.
• measurement: LinMeasurementLayerB, constraint: constructed with n2=self.n2 and hidden_units=hidden_units_meas.
• pr_assisted: PRAssistedLayer, constraint: constructed with length=self.n2, p_high, mode=sr_mode, resource_index, seed, and name="PRAssistedLayerB".
• combine: LinCombineLayerB, constraint: constructed with comms_size=self.comms_size and hidden_units=hidden_units_combine.
• Backward-compatible aliases: measure_layer, pr_layer, resource_layer, sr_layer, combine_layer referencing the corresponding primary layer attributes.

Planned (design-spec)

Not specified.

Deviations

Not specified.

Notes for Contributors

• compute_with_internal attempts two calling conventions for LinCombineLayerB: first positional (outcomes_b, comm_batch, training=...), then a dict {"outcomes": outcomes_b, "comm": comm_batch} if a TypeError occurs; maintain compatibility if evolving LinCombineLayerB.
• The "first_measurement" flag provided to PRAssistedLayer is always zeros with shape (B, 1); any changes to the PR-assisted contract should account for this fixed behavior.
• The call signature accepts **kwargs but does not use it; adding usage should be done carefully to preserve Keras compatibility.

Related

• Q_Sea_Battle.lin_measurement_layer_b.LinMeasurementLayerB
• Q_Sea_Battle.lin_combine_layer_b.LinCombineLayerB
• Q_Sea_Battle.pr_assisted_layer.PRAssistedLayer

Changelog

• 0.1: Initial documented version from module header.