Module train 

Training algorithms (PPO, DQN). Training algorithms (Burn backend).

Hosts the PPO and DQN trainers, the backend-agnostic optimizer abstraction, and the loss math shared between them. After phase 5 of the Burn migration (#82), Burn is the only tensor backend in the workspace.

Re-exports

pub use a2c::A2cConfig;

pub use a2c::A2cStats;

pub use a2c::A2cTrainer;

pub use a2c::compute_a2c_policy_loss;

pub use a2c::compute_a2c_value_loss;

pub use bc::BcConfig;

pub use bc::BcEpochStats;

pub use bc::BcTrainer;

pub use bc::Demonstrations;

pub use bc::compute_bc_loss;

pub use dqn::DQNConfig;

pub use dqn::DQNStepStatsBurn;

pub use dqn::DQNTrainerBurn;

pub use grad_clip::clip_grads_by_global_norm;

pub use grad_clip::global_grad_norm;

pub use optimizer::BackendOptimizer;

pub use optimizer::BurnOptimizer;

pub use ppo::AggregatedStats;

pub use ppo::AsyncActorLearnerConfig;

pub use ppo::PPOConfig;

pub use ppo::PPOTrainerBurn;

pub use ppo::TrainingStats;

pub use ppo::compute_entropy_loss;

pub use ppo::compute_policy_loss;

pub use ppo::compute_value_loss;

pub use ppo::generate_minibatch_indices;

pub use sac::SacConfig;

pub use sac::SacStepStats;

pub use sac::SacTrainer;

Modules

a2c

Synchronous Advantage Actor-Critic (A2C) trainer.

bc

Behavioral Cloning (BC) — supervised imitation learning.

dqn

DQN trainer (Burn backend).

grad_clip

Global gradient-norm clipping shared by the PPO trainers and the joint multi-agent trainer (issues #239, #299). Global gradient-norm clipping shared by the PPO trainers and the joint multi-agent trainer.

optimizer

Burn optimizer wrapper used by both PPO and DQN. Burn optimizer wrapper used by the PPO and DQN trainers.

ppo

PPO trainer (Burn backend).

sac

Soft Actor-Critic (SAC) trainer for continuous control.