Newer
Older
* Example solving the time dependent water flow 1D diffusion PDE:
* y_xx - y_t = 0, for (x, t) in [0, 1] x [0, 1]
* y(x, 0) = e^(-sqrt(0.5)x) * sin(-sqrt(0.5)x)
*
* -------------------------------------------
* Analytical solution:
* NN representation: sum over [a_i * (1 + e^(bi - x * w_ix - t * w_it))^(-1)]
* -------------------------------------------
* Optimal NN setting with biases (4 inner neurons)
* Path 1. wx = 0.51954589, wt = -0.48780445, b = 0.35656955, a = 1.69279158
* Path 2. wx = -1.24173503, wt = 1.13351300, b = 0.32528567, a = 1.69148458
* Path 3. wx = 0.64754127, wt = 0.95758760, b = -0.95852707, a = 2.77877453
* Path 4. wx = 1.65439557, wt = -0.31784248, b = -1.81237586, a = -3.96157108
* @author Michal Kravčenko
* @date 9.8.18
*/
#include <random>
#include <iostream>
#include <fstream>
#include "4neuro.h"
//y(x, t) = ... ai * f(wxi * x + wti * t - bi)
double eval_approx_y(double x, double t, size_t n_inner_neurons, std::vector<double> ¶meters){
double value= 0.0, wxi, wti, ai, bi, ei, ei1;
for(size_t i = 0; i < n_inner_neurons; ++i){
wxi = parameters[4 * i + 0];
wti = parameters[4 * i + 1];
ai = parameters[4 * i + 2];
bi = parameters[4 * i + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
value += ai / (ei1);
}
return value;
}
//yt(x, t) = ... (ai * wti * e^(bi - wti * t - wxi * x))/(e^(bi - wti * t - wxi * x) + 1)^2
double eval_approx_yt(double x, double t, size_t n_inner_neurons, std::vector<double> ¶meters){
double value= 0.0, wxi, wti, ai, bi, ei, ei1;
for(size_t i = 0; i < n_inner_neurons; ++i){
wxi = parameters[4 * i + 0];
wti = parameters[4 * i + 1];
ai = parameters[4 * i + 2];
bi = parameters[4 * i + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
value += ai * wti * ei / (ei1 * ei1);
}
return value;
}
//yx(x, t) = ... (ai * wxi * e^(bi - t * wti - wxi * x))/(e^(bi - t * wti - wxi * x) + 1)^2
double eval_approx_yx(double x, double t, size_t n_inner_neurons, std::vector<double> ¶meters){
double value= 0.0, wxi, wti, ai, bi, ei, ei1;
for(size_t i = 0; i < n_inner_neurons; ++i){
wxi = parameters[4 * i + 0];
wti = parameters[4 * i + 1];
ai = parameters[4 * i + 2];
bi = parameters[4 * i + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
value += (ai * wxi * ei1)/(ei1 * ei1);
}
return value;
}
//yxx(x, t) = ... (ai * wxi * e^(bi - t * wti - wxi * x))/(e^(bi - t * wti - wxi * x) + 1)^2
double eval_approx_yxx(double x, double t, size_t n_inner_neurons, std::vector<double> ¶meters){
double value= 0.0, wxi, wti, ai, bi, ei, ei1;
for(size_t i = 0; i < n_inner_neurons; ++i){
wxi = parameters[4 * i + 0];
wti = parameters[4 * i + 1];
ai = parameters[4 * i + 2];
bi = parameters[4 * i + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
value += (2 * ai * wxi * wxi * ei * ei) / (ei1 * ei1 * ei1) - (ai * wxi * wxi * ei) / (ei1 * ei1);
}
return value;
}
double eval_approx_da_y(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, bi, ei, ei1;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
return 1.0 / ei1;
}
double eval_approx_dwx_y(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
return (ai * x * ei)/(ei1 * ei1);
}
double eval_approx_dwt_y(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
return (ai * t * ei)/(ei1 * ei1);
}
double eval_approx_db_y(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, bi, ei, ai, ei1;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
return -(ai * ei)/(ei1 * ei1);
}
double eval_approx_da_yt(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, bi, ei, ei1;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
return (wti * ei)/(ei1 * ei1);
}
double eval_approx_dwx_yt(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
return (2 * ai * wti * x * ei * ei)/(ei1 * ei1 * ei1) - (ai * wti * x * ei)/(ei1 * ei1);
}
double eval_approx_dwt_yt(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
return -(ai * t * wti * ei) / (ei1 * ei1) + (2 * ai * t * wti * ei * ei)/(ei1 * ei1 * ei1) + (ai * ei)/(ei1 * ei1);
}
double eval_approx_db_yt(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ei1 = ei + 1.0;
return (ai * wti * ei) / (ei1 * ei1) - (2 * ai * wti * ei * ei) / (ei1 * ei1 * ei1);
}
double eval_approx_da_yxx(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1, ebp, eb, etx;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ebp= std::pow(l4n::E, bi + wxi * x + wti * t);
eb = std::pow(l4n::E, bi);
etx = std::pow(l4n::E, wxi * x + wti * t);
ei1 = eb + etx;
return -(wxi * wxi * ebp * (etx - eb))/(ei1 * ei1 * ei1);
}
double eval_approx_dwx_yxx(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1, ebp, eb, etx;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ebp= std::pow(l4n::E, bi + wxi * x + wti * t);
eb = std::pow(l4n::E, bi);
etx = std::pow(l4n::E, wxi * x + wti * t);
ei1 = eb + etx;
return (ai * wxi * wxi * x * ei) / ((ei + 1) * (ei + 1)) - (6 * ai * wxi * wxi * x * ei * ei) / ((ei + 1) * (ei + 1) * (ei + 1)) + (6 * ai * wxi *wxi * x * ei * ei * ei) / ((ei + 1) * (ei + 1) * (ei + 1) * (ei + 1)) - (2 * ai * wxi * ei) / ((ei + 1) * (ei + 1)) + (4 * ai * wxi * ei * ei)/((ei + 1) * (ei + 1) * (ei + 1));
}
double eval_approx_dwt_yxx(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1, ebp, eb, etx;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ebp= std::pow(l4n::E, bi + wxi * x + wti * t);
eb = std::pow(l4n::E, bi);
etx = std::pow(l4n::E, wxi * x + wti * t);
ei1 = eb + etx;
return (ai * t * wxi * wxi * ei) / ((ei + 1) * (ei + 1)) - (6 * ai * t * wxi * wxi * ei * ei) / ((ei + 1) * (ei + 1) * (ei + 1)) + (6 * ai * t * wxi * wxi * ei * ei * ei) / ((ei + 1) * (ei + 1) * (ei + 1) * (ei + 1));
}
double eval_approx_db_yxx(double x, double t, size_t neuron_idx, std::vector<double> ¶meters){
double wxi, wti, ai, bi, ei, ei1, ebp, eb, etx;
wxi = parameters[4 * neuron_idx + 0];
wti = parameters[4 * neuron_idx + 1];
ai = parameters[4 * neuron_idx + 2];
bi = parameters[4 * neuron_idx + 3];
Martin Beseda
committed
ei = std::pow(l4n::E, bi - wxi * x - wti * t);
ebp= std::pow(l4n::E, bi + wxi * x + wti * t);
eb = std::pow(l4n::E, bi);
etx = std::pow(l4n::E, wxi * x + wti * t);
ei1 = eb + etx;
return (ai * wxi * wxi * eb * ebp) / (ei1 * ei1 * ei1) - (ai * wxi * wxi * ebp * (etx - eb)) / (ei1 * ei1 * ei1) + (3 * ai * wxi * wxi * eb * ebp * (etx - eb)) / (ei1 * ei1 * ei1 * ei1);
}
void eval_step_size_simple(double &gamma, double val, double prev_val, double sk, double grad_norm, double grad_norm_prev){
if(val > prev_val){
gamma *= 0.99999;
}
if(sk <= 1e-3 || grad_norm < grad_norm_prev){
/* movement on a line */
/* new slope is less steep, speed up */
gamma *= 1.0005;
}
else if(grad_norm > grad_norm_prev){
/* new slope is more steep, slow down*/
gamma /= 1.0005;
}
else{
gamma /= 1.005;
}
// gamma *= 0.999999;
}
void eval_step_size_mk( double &gamma, double beta, double &c, double grad_norm_prev, double grad_norm, double fi, double fim ){
if( fi > fim )
{
c /= 1.0000005;
}
else if( fi < fim )
{
c *= 1.0000005;
}
gamma *= std::pow( c, 1.0 - 2.0 * beta) * std::pow( grad_norm_prev / grad_norm, 1.0 / c );
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
double calculate_gradient( std::vector<double> &data_points, size_t n_inner_neurons, std::vector<double> *parameters, std::vector<double> *gradient ){
size_t i, j, k;
double x, t, mem, derror, total_error, approx;
size_t train_size = data_points.size();
/* error of boundary condition: y(0, t) = sin(t) => e1 = 1/n * (sin(t) - y(0, t))^2 */
for(i = 0; i < train_size; ++i){
t = data_points[i];
mem = (std::sin(t) - eval_approx_y(0.0, t, n_inner_neurons, *parameters));
derror = 2.0 * mem / train_size;
for(j = 0; j < n_inner_neurons; ++j){
(*gradient)[4 * j + 0] -= derror * eval_approx_dwx_y(0, t, j, *parameters);
(*gradient)[4 * j + 1] -= derror * eval_approx_dwt_y(0, t, j, *parameters);
(*gradient)[4 * j + 2] -= derror * eval_approx_da_y(0, t, j, *parameters);
(*gradient)[4 * j + 3] -= derror * eval_approx_db_y(0, t, j, *parameters);
}
total_error += mem * mem / train_size;
}
/* error boundary condition: y(x, 0) = e^(-(0.5)^(0.5)x) * sin(-(0.5)^(0.5)x) => e2 = 1/n * (e^(-(0.5)^(0.5)x) * sin(-(0.5)^(0.5)x) - y(x, 0))^2 */
for(i = 0; i < train_size; ++i){
x = data_points[i];
Martin Beseda
committed
mem = (std::pow(l4n::E, -0.707106781 * x) * std::sin( -0.707106781 * x ) - eval_approx_y(x, 0.0, n_inner_neurons, *parameters));
derror = 2.0 * mem / train_size;
for(j = 0; j < n_inner_neurons; ++j){
(*gradient)[4 * j + 0] -= derror * eval_approx_dwx_y(x, 0, j, *parameters);
(*gradient)[4 * j + 1] -= derror * eval_approx_dwt_y(x, 0, j, *parameters);
(*gradient)[4 * j + 2] -= derror * eval_approx_da_y(x, 0, j, *parameters);
(*gradient)[4 * j + 3] -= derror * eval_approx_db_y(x, 0, j, *parameters);
}
total_error += mem * mem / train_size;
}
/* error of the governing equation: y_xx - y_t = 0 => e3 = 1/n^2 * (0 - y_xx + y_t)^2 */
for(i = 0; i < data_points.size(); ++i){
x = data_points[i];
for(j = 0; j < data_points.size(); ++j){
t = data_points[j];
approx = eval_approx_yxx(x, t, n_inner_neurons, *parameters) - eval_approx_yt(x, t, n_inner_neurons, *parameters);
mem = 0.0 - approx;
derror = 2.0 * mem / (train_size * train_size);
for(k = 0; k < n_inner_neurons; ++k){
(*gradient)[4 * k + 0] -= derror * (eval_approx_dwx_yxx(x, t, k, *parameters) - eval_approx_dwx_yt(x, t, k, *parameters));
(*gradient)[4 * k + 1] -= derror * (eval_approx_dwt_yxx(x, t, k, *parameters) - eval_approx_dwt_yt(x, t, k, *parameters));
(*gradient)[4 * k + 2] -= derror * ( eval_approx_da_yxx(x, t, k, *parameters) - eval_approx_da_yt(x, t, k, *parameters));
(*gradient)[4 * k + 3] -= derror * ( eval_approx_db_yxx(x, t, k, *parameters) - eval_approx_db_yt(x, t, k, *parameters));
}
total_error += mem * mem / (train_size * train_size);
}
}
return total_error;
void solve_example_gradient(std::vector<double> &guess, double accuracy, size_t n_inner_neurons, size_t train_size, double ds, double de, size_t n_test_points, double ts, double te){
Michal Kravcenko
committed
std::cout << "Finding a solution via a Gradient Descent method with adaptive step-length" << std::endl;
std::cout << "********************************************************************************************************************************************" <<std::endl;
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
/* SETUP OF THE TRAINING DATA */
std::vector<double> inp, out;
double frac, alpha, x;
double grad_norm = accuracy * 10.0, mem, ai, bi, wxi, wti, error, derror, approx, t, gamma, total_error, sk, sy, sx, sg, beta;
double grad_norm_prev = grad_norm;
size_t i, j, k, iter_idx = 0;
/* TRAIN DATA FOR THE GOVERNING DE */
std::vector<double> data_points(train_size);
/* ISOTROPIC TRAIN SET */
frac = (de - ds) / (train_size - 1);
for(i = 0; i < train_size; ++i){
data_points[i] = frac * i + ds;
// std::cout << data_points[i] << std::endl;
}
// /* CHEBYSCHEV TRAIN SET */
// alpha = PI / (train_size );
// frac = 0.5 * (de - ds);
// for(i = 0; i < train_size; ++i){
// x = (std::cos(PI - alpha * i) + 1.0) * frac + ds;
// data_points[i] = x;
// }
std::vector<double> *gradient_current = new std::vector<double>(4 * n_inner_neurons);
std::vector<double> *gradient_prev = new std::vector<double>(4 * n_inner_neurons);
std::vector<double> *params_current = new std::vector<double>(guess);
std::vector<double> *params_prev = new std::vector<double>(guess);
std::vector<double> *ptr_mem;
std::fill(gradient_current->begin(), gradient_current->end(), 0.0);
std::fill(gradient_prev->begin(), gradient_prev->end(), 0.0);
Michal Kravcenko
committed
// for (i = 0; i < n_inner_neurons; ++i) {
// wxi = (*params_current)[4 * i + 0];
// wti = (*params_current)[4 * i + 1];
// ai = (*params_current)[4 * i + 2];
// bi = (*params_current)[4 * i + 3];
//
// printf("Path %3d. wx = %15.8f, wt = %15.8f, b = %15.8f, a = %15.8f\n", (int)(i + 1), wxi, wti, bi, ai);
// }
gamma = 1.0;
double val = 0.0, prev_val, c = 2.0;
prev_val = 0.0;
while( grad_norm > accuracy) {
iter_idx++;
prev_val = val;
grad_norm_prev = grad_norm;
/* reset of the current gradient */
std::fill(gradient_current->begin(), gradient_current->end(), 0.0);
val = calculate_gradient( data_points, n_inner_neurons, params_current, gradient_current );
grad_norm = 0.0;
for(auto v: *gradient_current){
grad_norm += v * v;
}
grad_norm = std::sqrt(grad_norm);
/* Update of the parameters */
/* step length calculation */
if(iter_idx < 10 || iter_idx % 100 == 0){
/* fixed step length */
gamma = 0.1 * accuracy;
/* norm of the gradient calculation */
sk = 0.0;
for(i = 0; i < gradient_current->size(); ++i){
sx = (*gradient_current)[i] - (*gradient_prev)[i];
sk += sx * sx;
}
sk = std::sqrt(sk);
/* angle between two consecutive gradients */
double beta = 0.0, sx = 0.0;
for(i = 0; i < gradient_current->size(); ++i){
sx += (gradient_current->at( i ) * gradient_prev->at( i ));
sx /= grad_norm * grad_norm_prev;
Martin Beseda
committed
beta = std::sqrt(std::acos( sx ) / l4n::PI);
// eval_step_size_simple( gamma, val, prev_val, sk, grad_norm, grad_norm_prev );
eval_step_size_mk( gamma, beta, c, grad_norm_prev, grad_norm, val, prev_val );
}
for(i = 0; i < gradient_current->size(); ++i){
(*params_prev)[i] = (*params_current)[i] - gamma * (*gradient_current)[i];
}
/* switcheroo */
ptr_mem = gradient_prev;
gradient_prev = gradient_current;
gradient_current = ptr_mem;
ptr_mem = params_prev;
params_prev = params_current;
params_current = ptr_mem;
if(iter_idx % 1 == 0){
printf("Iteration %12d. Step size: %15.8f, C: %15.8f, Gradient norm: %15.8f. Total error: %10.8f\r", (int)iter_idx, gamma, c, grad_norm, val);
std::cout.flush();
}
}
printf("Iteration %12d. Step size: %15.8f, C: %15.8f, Gradient norm: %15.8f. Total error: %10.8f\r\n", (int)iter_idx, gamma, c, grad_norm, val);
std::cout.flush();
for (i = 0; i < n_inner_neurons; ++i) {
wxi = (*params_current)[4 * i + 0];
wti = (*params_current)[4 * i + 1];
ai = (*params_current)[4 * i + 2];
bi = (*params_current)[4 * i + 3];
Michal Kravcenko
committed
printf("Path %3d. wx%d = %15.8f, wt%d = %15.8f, b%d = %15.8f, a%d = %15.8f\n", (int)(i + 1), (int)(i + 1), wxi , (int)(i + 1), wti, (int)(i + 1), bi, (int)(i + 1), ai);
Michal Kravcenko
committed
std::cout << "********************************************************************************************************************************************" <<std::endl;
// for (i = 0; i < n_inner_neurons; ++i) {
// wxi = (*params_current)[4 * i + 0];
// wti = (*params_current)[4 * i + 1];
// ai = (*params_current)[4 * i + 2];
// bi = (*params_current)[4 * i + 3];
//
// printf("%f/(1+e^(%f-%fx-%ft)) + ", (int)(i + 1), ai, bi, wxi, wti);
// }
// printf("\n");
Michal Kravcenko
committed
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
// /* SOLUTION EXPORT */
// printf("Exporting file 'data_2d_pde1_y.txt' : %7.3f%%\r", 0.0);
// std::cout.flush();
//
// std::vector<double> input, output(1);
// std::ofstream ofs_y("data_2d_pde1_y.txt", std::ofstream::out);
// frac = (te - ts) / (n_test_points - 1);
// for(i = 0; i < n_test_points; ++i){
// x = i * frac + ts;
// for(j = 0; j < n_test_points; ++j){
// t = j * frac + ts;
// ofs_y << x << " " << t << " " << eval_approx_y(x, t, n_inner_neurons, *params_current) << std::endl;
// printf("Exporting file 'data_2d_pde1_y.txt' : %7.3f%%\r", (100.0 * (j + i * n_test_points)) / (n_test_points * n_test_points - 1));
// std::cout.flush();
// }
// }
// printf("Exporting file 'data_2d_pde1_y.txt' : %7.3f%%\n", 100.0);
// std::cout.flush();
// ofs_y.close();
//
// printf("Exporting file 'data_2d_pde1_yt.txt' : %7.3f%%\r", 0.0);
// std::cout.flush();
// std::ofstream ofs_t("data_2d_pde1_yt.txt", std::ofstream::out);
// frac = (te - ts) / (n_test_points - 1);
// for(i = 0; i < n_test_points; ++i){
// x = i * frac + ts;
// for(j = 0; j < n_test_points; ++j){
// t = j * frac + ts;
// ofs_t << x << " " << t << " " << eval_approx_yt(x, t, n_inner_neurons, *params_current) << std::endl;
// printf("Exporting file 'data_2d_pde1_yt.txt' : %7.3f%%\r", (100.0 * (j + i * n_test_points)) / (n_test_points * n_test_points - 1));
// std::cout.flush();
// }
// }
// printf("Exporting file 'data_2d_pde1_yt.txt' : %7.3f%%\n", 100.0);
// std::cout.flush();
// ofs_t.close();
//
// printf("Exporting file 'data_2d_pde1_yx.txt' : %7.3f%%\r", 0.0);
// std::cout.flush();
// std::ofstream ofs_x("data_2d_pde1_yx.txt", std::ofstream::out);
// frac = (te - ts) / (n_test_points - 1);
// for(i = 0; i < n_test_points; ++i){
// x = i * frac + ts;
// for(j = 0; j < n_test_points; ++j){
// t = j * frac + ts;
// ofs_x << x << " " << t << " " << eval_approx_yx(x, t, n_inner_neurons, *params_current) << std::endl;
// printf("Exporting file 'data_2d_pde1_yx.txt' : %7.3f%%\r", (100.0 * (j + i * n_test_points)) / (n_test_points * n_test_points - 1));
// std::cout.flush();
// }
// }
// printf("Exporting file 'data_2d_pde1_yx.txt' : %7.3f%%\n", 100.0);
// std::cout.flush();
// ofs_x.close();
//
// printf("Exporting file 'data_2d_pde1_yxx.txt' : %7.3f%%\r", 0.0);
// std::cout.flush();
// std::ofstream ofs_xx("data_2d_pde1_yxx.txt", std::ofstream::out);
// frac = (te - ts) / (n_test_points - 1);
// for(i = 0; i < n_test_points; ++i){
// x = i * frac + ts;
// for(j = 0; j < n_test_points; ++j){
// t = j * frac + ts;
// ofs_xx << x << " " << t << " " << eval_approx_yxx(x, t, n_inner_neurons, *params_current) << std::endl;
// printf("Exporting file 'data_2d_pde1_yxx.txt' : %7.3f%%\r", (100.0 * (j + i * n_test_points)) / (n_test_points * n_test_points - 1));
// std::cout.flush();
// }
// }
// printf("Exporting file 'data_2d_pde1_yxx.txt' : %7.3f%%\n", 100.0);
// std::cout.flush();
// ofs_xx.close();
//
// /* governing equation error */
// std::ofstream ofs_error("data_2d_pde1_first_equation_error.txt", std::ofstream::out);
// printf("Exporting file 'data_2d_pde1_first_equation_error.txt' : %7.3f%%\r", 0.0);
// for(i = 0; i < n_test_points; ++i){
// x = i * frac + ts;
// for(j = 0; j < n_test_points; ++j){
// t = j * frac + ts;
// ofs_error << x << " " << t << " " << std::fabs(eval_approx_yxx(x, t, n_inner_neurons, *params_current) - eval_approx_yt(x, t, n_inner_neurons, *params_current)) << std::endl;
// printf("Exporting file 'data_2d_pde1_first_equation_error.txt' : %7.3f%%\r", (100.0 * (j + i * n_test_points)) / (n_test_points * n_test_points - 1));
// std::cout.flush();
// }
// }
// printf("Exporting file 'data_2d_pde1_first_equation_error.txt' : %7.3f%%\n", 100.0);
// std::cout.flush();
// ofs_error.close();
//
// /* ISOTROPIC TEST SET FOR BOUNDARY CONDITIONS */
// /* first boundary condition & its error */
// std::ofstream ofs_bc_t("data_1d_pde1_yt.txt", std::ofstream::out);
// std::ofstream ofs_bc_x("data_1d_pde1_yx.txt", std::ofstream::out);
// printf("Exporting files 'data_1d_pde1_yt.txt' and 'data_1d_pde1_yx.txt' : %7.3f%%\r", 0.0);
// for(i = 0; i < n_test_points; ++i){
// x = frac * i + ts;
// t = frac * i + ts;
//
// double yt = std::sin(t);
Martin Beseda
committed
// double yx = std::pow(l4n::E, -0.707106781 * x) * std::sin( -0.707106781 * x );
Michal Kravcenko
committed
//
// double evalt = eval_approx_y(0, t, n_inner_neurons, *params_current);
// double evalx = eval_approx_y(x, 0, n_inner_neurons, *params_current);
//
// ofs_bc_t << i + 1 << " " << t << " " << yt << " " << evalt << " " << std::fabs(evalt - yt) << std::endl;
// ofs_bc_x << i + 1 << " " << x << " " << yx << " " << evalx << " " << std::fabs(evalx - yx) << std::endl;
//
// printf("Exporting files 'data_1d_pde1_yt.txt' and 'data_1d_pde1_yx.txt' : %7.3f%%\r", (100.0 * i) / (n_test_points - 1));
// std::cout.flush();
// }
// printf("Exporting files 'data_1d_pde1_yt.txt' and 'data_1d_pde1_yx.txt' : %7.3f%%\r", 100.0);
// std::cout.flush();
// ofs_bc_t.close();
// ofs_bc_x.close();
delete gradient_current;
delete gradient_prev;
delete params_current;
delete params_prev;
}
void solve_example_particle_swarm(double accuracy, size_t n_inner_neurons, size_t train_size, double ds, double de, size_t n_test_points, double ts, double te, size_t max_iters, size_t n_particles){
Michal Kravcenko
committed
std::cout << "Finding a solution via the Particle Swarm Optimization" << std::endl;
std::cout << "********************************************************************************************************************************************" <<std::endl;
/* solution properties */
/* do not change below */
size_t n_inputs = 2;
size_t n_equations = 3;
Martin Beseda
committed
l4n::DESolver solver_01( n_equations, n_inputs, n_inner_neurons );
Martin Beseda
committed
l4n::MultiIndex alpha_00( n_inputs );
l4n::MultiIndex alpha_01( n_inputs );
l4n::MultiIndex alpha_20( n_inputs );
alpha_00.set_partial_derivative(1, 0);
alpha_01.set_partial_derivative(1, 1);
alpha_20.set_partial_derivative(0, 2);
Michal Kravcenko
committed
solver_01.add_to_differential_equation( 0, alpha_20, "1.0" );
solver_01.add_to_differential_equation( 0, alpha_01, "-1.0" );
Michal Kravcenko
committed
solver_01.add_to_differential_equation( 1, alpha_00, "1.0" );
solver_01.add_to_differential_equation( 2, alpha_00, "1.0" );
//TODO neater data setup
std::vector<double> inp, out;
double frac, x, t;
/* TRAIN DATA FOR THE GOVERNING DE */
std::vector<double> test_bounds_2d = {ds, de, ds, de};
/* GOVERNING EQUATION RHS */
auto f1 = [](std::vector<double>&input) -> std::vector<double> {
std::vector<double> output(1);
output[0] = 0.0;
return output;
};
Martin Beseda
committed
l4n::DataSet ds_00(test_bounds_2d, train_size, f1, 1);
std::vector<std::pair<std::vector<double>, std::vector<double>>> data_vec_t;
std::vector<std::pair<std::vector<double>, std::vector<double>>> data_vec_x;
/* ISOTROPIC TRAIN SET */
frac = (de - ds) / (train_size - 1);
for(unsigned int i = 0; i < train_size; ++i){
inp = {0.0, frac * i};
out = {std::sin(inp[1])};
data_vec_t.emplace_back(std::make_pair(inp, out));
inp = {frac * i, 0.0};
Martin Beseda
committed
out = {std::pow(l4n::E, -0.707106781 * inp[0]) * std::sin( -0.707106781 * inp[0] )};
data_vec_x.emplace_back(std::make_pair(inp, out));
Martin Beseda
committed
l4n::DataSet ds_t(&data_vec_t);
l4n::DataSet ds_x(&data_vec_x);
Martin Beseda
committed
solver_01.set_error_function( 0, l4n::ErrorFunctionType::ErrorFuncMSE, &ds_00 );
solver_01.set_error_function( 1, l4n::ErrorFunctionType::ErrorFuncMSE, &ds_t );
solver_01.set_error_function( 2, l4n::ErrorFunctionType::ErrorFuncMSE, &ds_x );
Michal Kravcenko
committed
size_t total_dim = (2 + n_inputs) * n_inner_neurons;
/* PARTICLE SWARM TRAINING METHOD SETUP */
//must encapsulate each of the partial error functions
Michal Kravcenko
committed
std::vector<double> domain_bounds(2 * total_dim);
for(unsigned int i = 0; i < total_dim; ++i){
Michal Kravcenko
committed
domain_bounds[2 * i] = -10.0;
domain_bounds[2 * i + 1] = 10.0;
double c1 = 1.7, c2 = 1.7, w = 0.7;
double gamma = 0.5, epsilon = 0.02, delta = 0.9;
Michal Kravcenko
committed
solver_01.solve_via_particle_swarm( &domain_bounds, c1, c2, w, n_particles, max_iters, gamma, epsilon, delta );
size_t i, j;
std::vector<double> *w1_ptr = solver_01.get_solution( alpha_00 )->get_parameter_ptr_weights();
std::vector<double> *w2_ptr = solver_01.get_solution( alpha_00 )->get_parameter_ptr_biases();
Michal Kravcenko
committed
std::vector<double> export_params(total_dim);
Michal Kravcenko
committed
for(size_t i = 0; i < n_inner_neurons; ++i){
export_params[4 * i + 0] = w1_ptr->at(i);
export_params[4 * i + 1] = w1_ptr->at(n_inner_neurons + i);
export_params[4 * i + 2] = w1_ptr->at(2 * n_inner_neurons + i);
export_params[4 * i + 3] = w2_ptr->at( i );
Michal Kravcenko
committed
printf("Path %3d. wx%d = %15.8f, wt%d = %15.8f, b%d = %15.8f, a%d = %15.8f\n", (int)(i + 1), (int)(i + 1), export_params[4 * i + 0] , (int)(i + 1), export_params[4 * i + 1], (int)(i + 1), export_params[4 * i + 3], (int)(i + 1), export_params[4 * i + 2]);
Michal Kravcenko
committed
std::cout << "********************************************************************************************************************************************" << std::endl;
/* PRACTICAL END OF THE EXAMPLE */
Michal Kravcenko
committed
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
// /* SOLUTION EXPORT */
// for(i = 0; i < n_inner_neurons; ++i){
// export_params[4 * i + 0] = w1_ptr->at(i);
// export_params[4 * i + 1] = w1_ptr->at(n_inner_neurons + i);
// export_params[4 * i + 2] = w1_ptr->at(2 * n_inner_neurons + i);
// export_params[4 * i + 3] = w2_ptr->at( i );
// }
//
// printf("Exporting file 'data_2d_pde1_y.txt' : %7.3f%%\r", 0.0);
// std::cout.flush();
//
// std::vector<double> input, output(1);
// std::ofstream ofs("data_2d_pde1_y.txt", std::ofstream::out);
// frac = (te - ts) / (n_test_points - 1);
// for(i = 0; i < n_test_points; ++i){
// x = i * frac + ts;
// for(j = 0; j < n_test_points; ++j){
// t = j * frac + ts;
// ofs << x << " " << t << " " << eval_approx_y(x, t, n_inner_neurons, export_params) << std::endl;
// printf("Exporting file 'data_2d_pde1_y.txt' : %7.3f%%\r", (100.0 * (j + i * n_test_points)) / (n_test_points * n_test_points - 1));
// std::cout.flush();
// }
// }
// printf("Exporting file 'data_2d_pde1_y.txt' : %7.3f%%\n", 100.0);
// std::cout.flush();
// ofs.close();
//
// /* governing equation error */
// ofs = std::ofstream("data_2d_pde1_first_equation_error.txt", std::ofstream::out);
// printf("Exporting file 'data_2d_pde1_first_equation_error.txt' : %7.3f%%\r", 0.0);
// for(i = 0; i < n_test_points; ++i){
// x = i * frac + ts;
// for(j = 0; j < n_test_points; ++j){
// t = j * frac + ts;
// ofs << x << " " << t << " " << std::fabs(eval_approx_yxx(x, t, n_inner_neurons, export_params) - eval_approx_yt(x, t, n_inner_neurons, export_params)) << std::endl;
// printf("Exporting file 'data_2d_pde1_first_equation_error.txt' : %7.3f%%\r", (100.0 * (j + i * n_test_points)) / (n_test_points * n_test_points - 1));
// std::cout.flush();
// }
// }
// printf("Exporting file 'data_2d_pde1_first_equation_error.txt' : %7.3f%%\n", 100.0);
// std::cout.flush();
// ofs.close();
//
// /* ISOTROPIC TEST SET FOR BOUNDARY CONDITIONS */
// /* first boundary condition & its error */
// ofs = std::ofstream("data_1d_pde1_yt.txt", std::ofstream::out);
// std::ofstream ofs2("data_1d_pde1_yx.txt", std::ofstream::out);
// printf("Exporting files 'data_1d_pde1_yt.txt' and 'data_1d_pde1_yx.txt' : %7.3f%%\r", 0.0);
// for(i = 0; i < n_test_points; ++i){
// x = frac * i + ts;
// t = frac * i + ts;
//
// double yt = std::sin(t);
Martin Beseda
committed
// double yx = std::pow(l4n::E, -0.707106781 * x) * std::sin( -0.707106781 * x );
Michal Kravcenko
committed
//
// double evalt = eval_approx_y(0, t, n_inner_neurons, export_params);
// double evalx = eval_approx_y(x, 0, n_inner_neurons, export_params);
//
// ofs << i + 1 << " " << t << " " << yt << " " << evalt << " " << std::fabs(evalt - yt) << std::endl;
// ofs2 << i + 1 << " " << x << " " << yx << " " << evalx << " " << std::fabs(evalx - yx) << std::endl;
//
// printf("Exporting files 'data_1d_pde1_yt.txt' and 'data_1d_pde1_yx.txt' : %7.3f%%\r", (100.0 * i) / (n_test_points - 1));
// std::cout.flush();
// }
// printf("Exporting files 'data_1d_pde1_yt.txt' and 'data_1d_pde1_yx.txt' : %7.3f%%\r", 100.0);
// std::cout.flush();
// ofs2.close();
// ofs.close();
Michal Kravcenko
committed
std::cout << "Running lib4neuro Partial Differential Equation example 1" << std::endl;
std::cout << "********************************************************************************************************************************************" <<std::endl;
std::cout << " Governing equation: y_xx - y_t = 0, for (x, t) in [0, 1] x [0, 1]" << std::endl;
std::cout << "Dirichlet boundary condition: y(0, t) = sin(t), for t in [0, 1]" << std::endl;
std::cout << "Dirichlet boundary condition: y(x, 0) = exp(-sqrt(0.5)x) * sin(-sqrt(0.5)x), for x in [0, 1]" << std::endl;
std::cout << "********************************************************************************************************************************************" <<std::endl;
std::cout << "Expressing solution as y(x, t) = sum over [a_i / (1 + exp(bi - wxi*x - wti*t))], i in [1, n], where n is the number of hidden neurons" <<std::endl;
std::cout << "********************************************************************************************************************************************" <<std::endl;
unsigned int n_inner_neurons = 4;
unsigned int train_size = 10;
double ds = 0.0;
double de = 1.0;
unsigned int test_size = 100;
double ts = ds;
double te = de + 0;
size_t particle_swarm_max_iters = 500;
size_t n_particles = 10;
solve_example_particle_swarm(accuracy, n_inner_neurons, train_size, ds, de, test_size, ts, te, particle_swarm_max_iters, n_particles);
Michal Kravcenko
committed
// std::vector<double> init_guess(4 * n_inner_neurons);
// std::random_device seeder;
// std::mt19937 gen(seeder());
// std::uniform_real_distribution<double> dist(-1.0, 1.0);
// for(unsigned int i = 0; i < init_guess.size(); ++i){
// init_guess[i] = dist(gen);
// }
// init_guess = {-0.21709230, -0.26189447, 0.77853923, 0.41091127, -0.44311897, -0.99036349, 0.84912023, -0.16920743};
// solve_example_gradient(init_guess, accuracy, n_inner_neurons, train_size, ds, de, test_size, ts, te);