Examples with VectorPool nars.rl.horde.math.VectorPool used on opensource projects

Example 1 with VectorPool

use of nars.rl.horde.math.VectorPool in project narchy by automenta.

the class GQ method update.

public double update(RealVector x_t, double rho_t, double r_tp1, RealVector x_bar_tp1, double z_tp1) {
    if (x_t == null)
        return initEpisode();
    VectorPool pool = VectorPools.pool(x_t);
    delta_t = r_tp1 + beta_tp1 * z_tp1 + (1 - beta_tp1) * v.dotProduct(x_bar_tp1) - v.dotProduct(x_t);
    e.update((1 - beta_tp1) * lambda_t * rho_t, x_t);
    RealVector delta_e = pool.newVector(e.vect()).mapMultiplyToSelf(delta_t);
    ArrayRealVector tdCorrection = pool.newVector();
    if (x_bar_tp1 != null)
        tdCorrection.combineToSelf(0, 1, x_bar_tp1).mapMultiplyToSelf((1 - beta_tp1) * (1 - lambda_t) * e.vect().dotProduct(w));
    v.combineToSelf(1, alpha_v, pool.newVector(delta_e).combineToSelf(1, -1, tdCorrection));
    w.combineToSelf(1, alpha_w, delta_e.combineToSelf(1, -1, pool.newVector(x_t).mapMultiplyToSelf(w.dotProduct(x_t))));
    delta_e = null;
    pool.releaseAll();
    return delta_t;
}

Example 2 with VectorPool

use of nars.rl.horde.math.VectorPool in project narchy by automenta.

the class GreedyGQ method update.

public double update(RealVector x_t, A a_t, double r_tp1, double gamma_tp1, double z_tp1, RealVector x_tp1, A a_tp1) {
    rho_t = 0.0;
    if (a_t != null && x_t != null) /*!Vectors.isNull(x_t)*/
    {
        target.update(x_t);
        behaviour.update(x_t);
        rho_t = target.pi(a_t) / behaviour.pi(a_t);
    }
    // assert Utils.checkValue(rho_t);
    VectorPool pool = VectorPools.pool(prototype, gq.v.getDimension());
    RealVector sa_bar_tp1 = pool.newVector();
    // if (!Vectors.isNull(x_t) && !Vectors.isNull(x_tp1)) {
    if (x_t != null && x_tp1 != null) {
        target.update(x_tp1);
        for (A a : actions) {
            double pi = target.pi(a);
            if (pi == 0)
                continue;
            sa_bar_tp1.combineToSelf(1, pi, toStateAction.stateAction(x_tp1, a));
        }
    }
    RealVector phi_stat = x_t != null ? toStateAction.stateAction(x_t, a_t) : null;
    double delta_t = gq.update(phi_stat, rho_t, r_tp1, sa_bar_tp1, z_tp1);
    pool.releaseAll();
    return delta_t;
}

Example 3 with VectorPool

use of nars.rl.horde.math.VectorPool in project narchy by automenta.

the class GTDLambda method update.

@Override
public double update(double pi_t, double b_t, RealVector x_t, RealVector x_tp1, double r_tp1, double gamma_tp1, double z_tp1) {
    if (x_t == null)
        return initEpisode(gamma_tp1);
    VectorPool pool = VectorPools.pool(e.vect());
    v_t = v.dotProduct(x_t);
    delta_t = r_tp1 + (1 - gamma_tp1) * z_tp1 + gamma_tp1 * v.dotProduct(x_tp1) - v_t;
    // Update traces
    e.update(gamma_t * lambda, x_t);
    double rho_t = pi_t / b_t;
    e.vect().mapMultiplyToSelf(rho_t);
    // Compute correction
    ArrayRealVector correctionVector = pool.newVector();
    if (x_tp1 != null) {
        correction = e.vect().dotProduct(w);
        correctionVector.combineToSelf(1, correction * gamma_tp1 * (1 - lambda), x_tp1);
    }
    // Update parameters
    RealVector deltaE = pool.newVector(e.vect()).mapMultiplyToSelf(delta_t);
    v.combineToSelf(1, alpha_v, pool.newVector(deltaE).combineToSelf(1, -1, correctionVector));
    w.combineToSelf(1, alpha_w, deltaE.combineToSelf(1, -w.dotProduct(x_t), x_t));
    deltaE = null;
    gamma_t = gamma_tp1;
    pool.releaseAll();
    return delta_t;
}