Crate pc_rl_core 

Predictive Coding Actor-Critic framework.

A publishable reinforcement learning library implementing a novel architecture where the actor uses iterative top-down/bottom-up predictive coding inference loops instead of standard feedforward passes, combined with a standard MLP critic for value estimation.

Key Components

• PcActor — Predictive coding network with variable hidden topology, iterative inference loop, and surprise scoring.
• MlpCritic — Standard MLP value function with MSE loss backpropagation.
• PcActorCritic — Integrated agent: act, learn (episodic/continuous), surprise-based scheduling, save/load.
• serializer — JSON weight persistence with checkpointing support.

Re-exports

pub use activation::Activation;

pub use error::PcError;

pub use layer::Layer;

pub use layer::LayerDef;

pub use linalg::cpu::CpuLinAlg;

pub use linalg::golub_kahan::GolubKahanSvd;

pub use linalg::golub_kahan::SvdError;

pub use linalg::LinAlg;

pub use matrix::argmax_masked;

pub use matrix::cca_neuron_alignment;

pub use matrix::rms_error;

pub use matrix::sample_from_probs;

pub use matrix::softmax_masked;

pub use matrix::Matrix;

pub use matrix::GRAD_CLIP;

pub use matrix::WEIGHT_CLIP;

pub use mlp_critic::MlpCritic;

pub use mlp_critic::MlpCriticConfig;

pub use mlp_critic::MlpCriticWeights;

pub use pc_actor::InferResult;

pub use pc_actor::PcActor;

pub use pc_actor::PcActorConfig;

pub use pc_actor::SelectionMode;

pub use pc_actor_critic::ActivationCache;

pub use pc_actor_critic::PcActorCritic;

pub use pc_actor_critic::PcActorCriticConfig;

pub use pc_actor_critic::TrajectoryStep;

pub use serializer::checkpoint_filename;

pub use serializer::load_agent;

pub use serializer::load_agent_generic;

pub use serializer::save_agent;

pub use serializer::save_checkpoint;

pub use serializer::AgentMetadata;

pub use serializer::PcActorWeights;

pub use serializer::SaveFile;

pub use serializer::TrainingMetrics;

Modules

activation

Activation functions for neural network layers.

error

Crate-wide error type for pc_rl_core.

layer

Dense neural network layer.

linalg

Linear algebra trait abstraction for backend-agnostic compute.

matrix

Dense matrix operations and vector utilities for neural networks.

mlp_critic

Standard MLP critic (value function) for the PC Actor-Critic agent.

pc_actor

Predictive Coding Actor Network.

pc_actor_critic

Integrated PC Actor-Critic agent.

serializer

JSON-based weight persistence for the PC-Actor-Critic agent.

Type Aliases

LayerCpu

Type alias: CPU-backed layer.

MlpCriticCpu

Type alias: CPU-backed MLP critic.

PcActorCpu

Type alias: CPU-backed PC actor.

PcActorCriticCpu

Type alias: CPU-backed PC actor-critic agent.