Struct zero_sum::impls::tak::evaluator::AnnEvaluator

pub struct AnnEvaluator { /* fields omitted */ }

Uses an artificial neural network to evaluate the tak state. The network has three hidden layers using ReLU activation, and uses TanH activation on the output. It uses ADADELTA to perform gradient descent.

This is largely an experimental evaluator. While it has demonstrated its potential as a stronger evaluator of tak positions than the StaticEvaluator, it is an order of magnitude slower, making it unfit for use in a real-time game.

Methods

impl AnnEvaluator

fn new() -> AnnEvaluator

Creates a new evaluator, randomly initializing the network.

fn from_file(filename: &str) -> Result<AnnEvaluator, String>

Loads in a network state previously serialized with the to_file method.

fn to_file(&self, filename: &str)

Writes the current network state to a file.

fn train_batch(
    &mut self,
    positions: &[State],
    labels: &[Evaluation],
    error: Option<&mut f32>
)

Trains the network on positions, against labels. Optionally will return the average amount of error per input in error.

fn train_batch_tdleaf(
    &mut self,
    positions: &[State],
    error: Option<&mut f32>,
    thread_count: usize
)

Use temporal difference learning (TD-Leaf algorithm) to train the system through self-play. positions are used as starting points for self-play. Optionally returns the average amount of error per input in error.

Trait Implementations

impl Clone for AnnEvaluator

fn clone(&self) -> AnnEvaluator

Returns a copy of the value.

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

Performs copy-assignment from source.

impl Evaluator for AnnEvaluator

type State = State

type Evaluation = Evaluation

fn evaluate(&self, state: &State) -> Evaluation

Returns the evaluation of state.

fn evaluate_plies(
    &self,
    state: &Self::State,
    plies: &[<Self::State as State>::Ply]
) -> Self::Evaluation

Returns the evaluation of state after executing plies.