Struct border_async_trainer::Actor

pub struct Actor<A, E, P, R>where
    A: Agent<E, R> + SyncModel,
    E: Env,
    P: StepProcessorBase<E>,
    R: ReplayBufferBase<PushedItem = P::Output>,{ /* private fields */ }

Runs interaction between an Agent and an Env, then generates transitions.

flowchart TB E["Agent
(in AsyncTrainer)"]-->|SyncModel::ModelInfo|A subgraph D[Actor] A[Agent]-->|Env::Act|B[Env] B-->|Env::Obs|A B-->|Step<E: Env>|C[StepProcessorBase] end C-->|ReplayBufferBase::PushedItem|F[ReplayBufferProxy]

This diagram shows interaction of Agent, Env and StepProcessorBase, as shown in border_core::Trainer. However, this diagram also shows that The Agent in the Actor periodically synchronizes with the Agent in AsyncTrainer via SyncModel::ModelInfo, and the transitions generated by StepProcessorBase are sent to the ReplayBufferProxy.

See also the diagram in AsyncTrainer.

Implementations

impl<A, E, P, R> Actor<A, E, P, R>where A: Agent<E, R> + SyncModel, E: Env, P: StepProcessorBase<E>, R: ReplayBufferBase<PushedItem = P::Output>,

pub fn build( id: usize, agent_config: A::Config, env_config: E::Config, step_proc_config: P::Config, replay_buffer_config: ReplayBufferProxyConfig, stop: Arc<Mutex<bool>>, env_seed: i64, stats: Arc<Mutex<Option<ActorStat>>> ) -> Self

pub fn run( &mut self, sender: Sender<PushedItemMessage<R::PushedItem>>, model_info: Arc<Mutex<(usize, A::ModelInfo)>>, guard: Arc<Mutex<bool>>, guard_init_model: Arc<Mutex<bool>> )

Runs sampling loop until self.stop becomes true.

When finishes, this method sets ActorStat.