void rl_nac::update( double reward, double *statefeats ) {
vector_t grad_guess;
// incorporate new information
add_obs( reward, statefeats );
// we don't want to solve the linear system of equations every time.
if ( time_to_check() ) {
if ( inject_delay ) {
//cout << "injecting delay of " << injection_nanos << ", for rl_to_sleepidle_ratio = " << rl_to_sleepidle_ratio << endl;
slowdown_part1();
}
solve_for_grad( grad_guess ); //put result in grad_guess
if ( check_for_convergence(grad_guess, prev_grad_guess) ) {
farm_out_grad( grad_guess );
take_gradient_step();
prev_grad_guess.Fill( 0 );
} else {
prev_grad_guess = grad_guess;
}
if ( inject_delay )
slowdown_part2();
}
for ( int a=0; a<act_cnt; a++ ) {
// drive parameters towards zero
Add( -PARAM_REGULARIZER, (*acts)[a]->params, (*acts)[a]->params );
}
}
bool affine_propagator::process_all_eqs(const affine r[], const affine vars[], double& maxProgress) {
for (int p=0; p<n_vars; ++p) {
bool has_converged = check_for_convergence(vars);
if (has_converged) {
maxProgress = 2.0;
return false;
}
const affine& aa = r[p];
//cout << scientific;
//cout << endl << "Eq. #" << p << endl << aa << endl;
const bool to_discard = process_vars_in_eq(aa, vars);
if (to_discard) {
affine::isValid = false;
return true;
}
if (!check_for_zero(r, p, p)) {
//cout << "Deleted" << endl;
affine::isValid = false;
return true;
}
}
return false;
}
