Struct DqnConfig 

pub struct DqnConfig {
    pub gamma: f64,
    pub learning_rate: f64,
    pub batch_size: usize,
    pub buffer_capacity: usize,
    pub min_replay_size: usize,
    pub target_update_freq: usize,
    pub hidden_sizes: Vec<usize>,
    pub epsilon_start: f64,
    pub epsilon_end: f64,
    pub epsilon_decay_steps: usize,
}

Configuration for a DQN agent.

All hyperparameters live here. Pass this to DqnAgent::new(). The defaults reflect standard DQN practice suitable for moderately complex environments. Simple environments like CartPole will want smaller buffer/warmup values and faster epsilon decay.

Fields

gamma: f64

Discount factor γ. Controls how much future rewards are valued. Typical values: 0.95–0.999. Default: 0.99

learning_rate: f64

Learning rate for the Adam optimiser. Default: 1e-4

batch_size: usize

Number of experiences sampled per gradient update. Default: 32

buffer_capacity: usize

Maximum number of experiences in the replay buffer. Oldest are overwritten when full. Default: 100_000

min_replay_size: usize

Minimum number of experiences collected before training begins. During warm-up, actions are sampled randomly. Must be >= batch_size. Default: 10_000

target_update_freq: usize

Number of steps between hard target network updates.

The target network is a frozen copy of the online network used to compute stable TD targets. Updating it too frequently causes instability; too rarely slows learning. Default: 1_000

hidden_sizes: Vec<usize>

Hidden layer sizes for the Q-network.

The network architecture is: obs_size -> hidden[0] -> hidden[1] -> ... -> num_actions All hidden layers use ReLU activations. Default: [128, 128]

epsilon_start: f64

Starting epsilon for ε-greedy exploration. At step 0, actions are random with this probability. Default: 1.0

epsilon_end: f64

Final epsilon after decay is complete. Default: 0.05

epsilon_decay_steps: usize

Number of steps over which epsilon decays linearly from epsilon_start to epsilon_end. Default: 50_000

Implementations

impl DqnConfig

pub fn epsilon_at(&self, step: usize) -> f64

Compute the current epsilon given the number of elapsed steps.

Decays linearly from epsilon_start to epsilon_end over epsilon_decay_steps, then stays flat.

Trait Implementations

impl Clone for DqnConfig

fn clone(&self) -> DqnConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DqnConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for DqnConfig

fn default() -> Self

Returns the “default value” for a type.