Struct DefaultEvaluator

pub struct DefaultEvaluator<E: Env> { /* private fields */ }

A default implementation of the Evaluator trait.

This evaluator runs a specified number of episodes and calculates the average return (cumulative reward) across all episodes. It is useful for:

• Evaluating the performance of trained agents
• Comparing different policies or algorithms
• Monitoring training progress

Type Parameters

• E - The environment type

Examples

let config = EnvConfig::default();
let mut evaluator = DefaultEvaluator::new(&config, 42, 10)?;

// Evaluate a policy
let record = evaluator.evaluate(&mut agent)?;
println!("Average return: {}", record.get_scalar("Episode return")?);

Implementations

impl<E: Env> DefaultEvaluator<E>

pub fn new(config: &E::Config, seed: i64, n_episodes: usize) -> Result<Self>

Constructs a new DefaultEvaluator.

Arguments

• config - Configuration for the environment
• seed - Random seed for environment initialization
• n_episodes - Number of episodes to run during evaluation

Returns

A new evaluator instance

Examples

let config = EnvConfig::default();
let evaluator = DefaultEvaluator::new(&config, 42, 10)?;

Trait Implementations

impl<E: Env> Evaluator<E> for DefaultEvaluator<E>

fn evaluate<R>( &mut self, policy: &mut Box<dyn Agent<E, R>>, ) -> Result<(f32, Record)>

where R: ReplayBufferBase,

Evaluates a policy by running multiple episodes and calculating the average return.

This method:

1. Runs the specified number of episodes
2. For each episode:
   • Resets the environment with a unique index
   • Runs the episode until termination
   • Accumulates the total reward
3. Returns the average return across all episodes

Arguments

• policy - The policy to evaluate

Returns

A tuple of (performance metric, Record containing the evaluation results)

Errors

Returns an error if:

• The environment fails to reset
• The environment fails to step