Struct ark_api::ml::Experiment

pub struct Experiment { /* private fields */ }

A machine learning context. Takes batches of observations, and produces corresponding batches of actions.

Implementations

impl Experiment

pub fn new( host: &HiveHost, game_name: &str, experiment_name: &str, num_remote_workers: u32, config: &str, checkpoint: Option<&str>, training_duration_in_seconds: u64, protocol: &ProtocolConfig ) -> impl Future<Output = Result<Self, Error>>

Start training with a new experiment.

host where the hive server exists. game_name identifier for a type of game, e.g. ‘pong’ or ‘moreau-arena’. experiment_name identifier for this experiment. This can be any arbitrary data. If the server already has an experiment with that name then a number will be appended to make it unique. num_features number of floating point values the neural network takes as input in each observation. num_actions number of floating point values the neural network outputs for each observation. num_remote_workers number of workers that should be started in the cloud. Using a non-zero number is only supported when using a cloud host. config is a piece of data that contains all the necessary information to re-create the training run. This is not used anywhere internally, but when connecting to an existing experiment you can query it using the Experiment::config method. The only constraint on this data is that it must be valid json. This is very useful for remote workers as they can download this configuration. checkpoint optional reference to checkpoint data to load from. TODO: Api for listing checkpoints. training_duration_in_seconds Number of seconds that the training will run for. The hive server will stop the experiment after this amount of time. This is primarily useful when training in the cloud with remote workers,

pub async fn new_from_settings( host: HiveHost, settings: StartExperimentSettings ) -> Result<Self, Error>

Start training with a new experiment.

host where the hive server exists. settings all data required to spawn the experiment

pub fn connect_to_experiment( host: &HiveHost, run_id: &str ) -> impl Future<Output = Result<Self, Error>>

Connects to an already running experiment

pub fn config(&self) -> impl Future<Output = Result<ExperimentConfig, Error>>

Retrieves the configuration for the experiment

This may require a download from the hive server.

pub fn raw_config(&self) -> impl Future<Output = Result<String, Error>>

Retrieves the configuration for the experiment

This may require a download from the hive server.

pub fn metrics(&self) -> impl Future<Output = Result<Vec<(String, f32)>, Error>>

Retrieves current training metrics for Hive

pub fn snapshot(&self) -> impl Future<Output = Result<Vec<u8>, Error>>

Retrieves a trained snapshot.

pub fn push_training_experiences( &self, observations: &[Observation<'_>] ) -> Result<Vec<Response>, Error>

Trains using the current batch of observations.

pub fn push_demonstration_experiences( &self, demonstrations: &[Demonstration<'_>] ) -> Result<(), Error>

Send demonstration experiences recorded in ark to Hive. Will return an error if the Hive protocol doesn’t support imitation learning

pub fn push_augmented_experiences( &self, observations: &[AugmentedObservation<'_>] ) -> Result<(), Error>

Send augmented experiences augmented in ark to Hive.

The list of observations represents one timestep for a set of agents. You should not include multiple observations for a single agent, call this method multiple times instead.

pub fn stop_experiment(self) -> Result<(), Error>

Stops an experiment and all associated workers.

If you just want to disconnect this client (without stopping potential remote workers) then simply drop the experiment.

Trait Implementations

impl Drop for Experiment

fn drop(&mut self)

Executes the destructor for this type.