Struct DqnPolicy 

pub struct DqnPolicy<E, Enc, Act, B: Backend> { /* private fields */ }

A DQN agent in inference-only mode.

Holds just the Q-network, encoder, and action mapper – no optimizer, no replay buffer, no exploration. Requires only B: Backend (not AutodiffBackend), so it can run on plain NdArray without any autodiff overhead.

Load from a checkpoint saved by DqnAgent::save:

use burn::backend::NdArray;

let policy = DqnPolicy::<CartPoleEnv, _, _, NdArray>::new(
    VecEncoder::new(4),
    UsizeActionMapper::new(2),
    &config,
    device,
)
.load("cartpole_dqn")?;

Implementations

impl<E, Enc, Act, B> DqnPolicy<E, Enc, Act, B>

where E: Environment, Enc: ObservationEncoder<E::Observation, B>, Act: DiscreteActionMapper<E::Action>, B: Backend,

pub fn new( encoder: Enc, action_mapper: Act, config: &DqnConfig, device: B::Device, ) -> Self

Build an uninitialised policy with the given architecture.

The network weights are random until load is called.

pub fn from_network( net: QNetwork<B>, encoder: Enc, action_mapper: Act, device: B::Device, ) -> Self

Create a policy directly from a pre-built network.

Used by DqnAgent::into_policy() to convert a trained agent into an inference-only policy without touching the disk.

pub fn load(self, path: impl AsRef<Path>) -> Result<Self, RecorderError>

Load network weights from a checkpoint file.

The checkpoint must have been saved with DqnAgent::save (.mpk format). The architecture (hidden sizes) must match exactly.

Trait Implementations

impl<E, Enc, Act, B> Policy<<E as Environment>::Observation, <E as Environment>::Action> for DqnPolicy<E, Enc, Act, B>

where E: Environment, Enc: ObservationEncoder<E::Observation, B>, Act: DiscreteActionMapper<E::Action>, B: Backend,

fn act(&self, obs: &E::Observation) -> E::Action

Select an action given the current observation.