Struct PcActor 

pub struct PcActor<L: LinAlg = CpuLinAlg> {
    pub config: PcActorConfig,
    /* private fields */
}

Predictive coding actor network.

Uses iterative top-down/bottom-up inference loops to produce stable hidden representations and output logits.

Generic over a LinAlg backend L. Defaults to CpuLinAlg.

Examples

use pc_rl_core::activation::Activation;
use pc_rl_core::layer::LayerDef;
use pc_rl_core::linalg::cpu::CpuLinAlg;
use pc_rl_core::pc_actor::{PcActor, PcActorConfig, SelectionMode};
use rand::SeedableRng;
use rand::rngs::StdRng;

let config = PcActorConfig {
    input_size: 9,
    hidden_layers: vec![LayerDef { size: 18, activation: Activation::Tanh }],
    output_size: 9,
    output_activation: Activation::Tanh,
    alpha: 0.1, tol: 0.01, min_steps: 1, max_steps: 20,
    lr_weights: 0.01, synchronous: true, temperature: 1.0,
    local_lambda: 1.0,
    residual: false,
    rezero_init: 0.001,
};
let mut rng = StdRng::seed_from_u64(42);
let actor: PcActor = PcActor::new(CpuLinAlg::new(), config, &mut rng).unwrap();
let result = actor.infer(&[0.0; 9]);
assert_eq!(result.y_conv.len(), 9);

Fields

config: PcActorConfig

Actor configuration.

Implementations

impl<L: LinAlg> PcActor<L>

pub fn new( backend: L, config: PcActorConfig, rng: &mut impl Rng, ) -> Result<Self, PcError>

Creates a new PC actor with Xavier-initialized layers.

Arguments

• config - Actor configuration specifying topology and hyperparameters.
• rng - Random number generator for weight initialization.

Errors

Returns PcError::ConfigValidation if input_size, output_size, or temperature are invalid.

pub fn crossover( parent_a: &PcActor<L>, parent_b: &PcActor<L>, caches_a: &[L::Matrix], caches_b: &[L::Matrix], alpha: f64, child_config: PcActorConfig, rng: &mut impl Rng, ) -> Result<Self, PcError>

Creates a child actor by crossing over two parent actors using CCA neuron alignment.

Aligns hidden neurons functionally via CCA before blending weights. Input and output layers use positional crossover (no permutation problem).

Arguments

• parent_a - First parent (reference, typically higher fitness).
• parent_b - Second parent (aligned to A via CCA).
• caches_a - Per-layer activation matrices for parent A [batch × neurons].
• caches_b - Per-layer activation matrices for parent B [batch × neurons].
• alpha - Blending weight: 1.0 = all A, 0.0 = all B.
• child_config - Topology configuration for the child network.
• rng - Random number generator for Xavier initialization.

Errors

Returns PcError::ConfigValidation if child_config is invalid.

pub fn latent_size(&self) -> usize

Returns the total size of the latent concatenation (sum of hidden layer sizes).

pub fn infer(&self, input: &[f64]) -> InferResult<L>

pub fn select_action( &self, y_conv: &L::Vector, valid_actions: &[usize], mode: SelectionMode, rng: &mut impl Rng, ) -> usize

Selects an action given converged output logits and valid actions.

Arguments

• y_conv - Output logits from inference.
• valid_actions - Indices of valid actions.
• mode - Training (stochastic) or Play (deterministic).
• rng - Random number generator (used only in Training mode).

Panics

Panics if valid_actions is empty.

pub fn update_weights( &mut self, output_delta: &[f64], infer_result: &InferResult<L>, input: &[f64], surprise_scale: f64, )

Updates network weights using a blend of backprop and local PC error.

The local_lambda config controls the blend: 1.0 = pure backprop, 0.0 = pure local PC learning (Millidge et al. 2022), intermediate = hybrid.

Arguments

• output_delta - Error signal at the output layer.
• infer_result - Result from the most recent inference.
• input - Original input that was fed to infer.
• surprise_scale - Multiplier on learning rate based on surprise.

Panics

Panics if input.len() != config.input_size.

pub fn to_weights(&self) -> PcActorWeights

Extracts a serializable snapshot of current weights.

Converts generic layers and skip projections to CPU-backed types.

pub fn from_weights( backend: L, config: PcActorConfig, weights: PcActorWeights, ) -> Result<Self, PcError>

Restores an actor from saved weights without requiring an RNG.

Converts CPU-backed weight snapshots to the target backend L. Validates that all weight matrix dimensions and bias lengths match the expected topology from config.

Errors

Returns PcError::DimensionMismatch if any weight matrix or bias vector has dimensions inconsistent with the config topology.

Trait Implementations

impl<L: Debug + LinAlg> Debug for PcActor<L>

where L::Matrix: Debug,

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.