source: liacs/ai/poker/nn.c@ 47

/*
 * Rick van der Zwet
 * 0433373
 * OS Assigment 3
 * Licence: BSD
 * $Id: nn.c 557 2008-04-08 22:57:09Z rick $
*/

#include <sysexits.h>
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <time.h>

/* NOTE: All first knobs are bias knobs or hidden stale knobs
 * - Validation is done using rounding, please make outputs discrete or
 * alter validation function
 */

/* Allow uniform and easy calls at functions */
#define TRUE 1
#define FALSE 0


/* Network variables */
/*NOTE: first node is 'hidden' bias knob */
#ifndef INPUT_SIZE
#define INPUT_SIZE 11
#endif

/*NOTE: first node is 'hidden' bias knob */
#ifndef HIDDEN_SIZE
#define HIDDEN_SIZE 11
#endif

/*NOTE: first node is 'hidden' 'lame' knob */
#ifndef OUTPUT_SIZE
#define OUTPUT_SIZE 11
#endif

/* Learn speed alpha of network */
#ifndef LEARN_SPEED
#define LEARN_SPEED 0.5
#endif

/* After QUALITY_ROUND trainingset check quality of network */
#define QUALITY_ROUND 100

/* Training set, to be used to train network */
char * file_training = "data/training.txt";
/* Validation set, to be used to test end result of network */
char * file_validate = "data/validate.txt";
/* Quality set, to be used to do quick testing whether network is
 * improving
 */
char * file_quality = "data/quality.txt";

/* Globally defined arrays, which represent the network */
double hidden[HIDDEN_SIZE];
double input[INPUT_SIZE];
double output[OUTPUT_SIZE];
double target[OUTPUT_SIZE];
double weight_HtoO[HIDDEN_SIZE][OUTPUT_SIZE];
double weight_ItoH[INPUT_SIZE][HIDDEN_SIZE];

#define WEIGHT_NOT_USED -99999

void stdInit() {
    int i;
    /* Should never change, been using */
    for (i = 0; i < INPUT_SIZE; i++)
        weight_ItoH[i][0] = WEIGHT_NOT_USED;
    for (i = 0; i < HIDDEN_SIZE; i++)
        weight_HtoO[i][0] = WEIGHT_NOT_USED;
}


/* Random init of weights */
void randInit() {
    int i,j;
    
    /* Different numbers every call */
    srandom(time(NULL));

    for (i = 0; i < INPUT_SIZE; i++)
        for ( j = 1; j < HIDDEN_SIZE; j++) {
            weight_ItoH[i][j] = (double)(random() % 100) / 100;
        }

    for (i = 0; i < HIDDEN_SIZE; i++)
        for (j = 1; j < OUTPUT_SIZE; j++)
            weight_HtoO[i][j] = (double)(random() % 100) / 100;

    stdInit();
}

/* Fixed init of weights */
void fixedInit() {
    int i,j;
    for (i = 0; i < INPUT_SIZE; i++)
        for ( j = 1; j < HIDDEN_SIZE; j++) {
            weight_ItoH[i][j] = 0.5;
        }

    for (i = 0; i < HIDDEN_SIZE; i++)
        for (j = 1; j < OUTPUT_SIZE; j++)
            weight_HtoO[i][j] = 0.5;

    stdInit();
}

/* Define exact wights, used for debugging calculations
 * other flags INPUT = 2, HIDDEN = 2, OUTPUT = 1
 */
void debugInit() {
    stdInit();
    weight_ItoH[0][1] = 1;
    weight_ItoH[0][2] = 1;
    weight_ItoH[1][1] = 0.62;
    weight_ItoH[1][2] = 0.42;
    weight_ItoH[2][1] = 0.55;
    weight_ItoH[2][2] = -0.17;

    weight_HtoO[0][1] = 1;
    weight_HtoO[1][1] = 0.35;
    weight_HtoO[2][1] = 0.81;
}

/*  calculate Aj's and Ai's (outputs) */
void nnCalc() {
    int i,j;
    double total;
    for (i = 1; i < HIDDEN_SIZE; i++) {
        total = 0;
        for (j = 0; j < INPUT_SIZE; j++)
            total += weight_ItoH[j][i] * input[j];
        hidden[i] = 1 / ( 1 + exp(total * (-1)));
    }

    for (i = 1; i < OUTPUT_SIZE; i++) {
        total = 0;
        for (j = 0; j < HIDDEN_SIZE; j++)
            total += weight_HtoO[j][i] * hidden[j];
        output[i] = 1 / ( 1 + exp(total * (-1)));
    }

}

/* train network, NOTE: nnCalc needs to be called first */
void nnTrain() {
    int i,j;
    double hidden_delta[HIDDEN_SIZE];
    double output_delta[OUTPUT_SIZE];
    double output_error[OUTPUT_SIZE];
    double hidden_sum_delta[HIDDEN_SIZE];

    for (i = 1; i < OUTPUT_SIZE; i++) {
        output_error[i] = target[i] - output[i];
        output_delta[i] = output_error[i] * output[i] * (1 - output[i]);
    }

    for (i = 0; i < HIDDEN_SIZE; i++) {
        hidden_sum_delta[i] = 0;
        for (j = 1; j < OUTPUT_SIZE; j++)
            hidden_sum_delta[i] += weight_HtoO[i][j] * output_delta[j];
        hidden_delta[i] = hidden[i] * (1 - hidden[i]) *
        hidden_sum_delta[i];
     }

    for (i = 0; i < HIDDEN_SIZE; i++)
        for (j = 1; j < OUTPUT_SIZE; j++) {
            weight_HtoO[i][j] = weight_HtoO[i][j] + LEARN_SPEED *
            hidden[i] * output_delta[j];
        }

    for (i = 0; i < INPUT_SIZE; i++)
        for (j = 1; j < HIDDEN_SIZE; j++) {
            weight_ItoH[i][j] = weight_ItoH[i][j] + LEARN_SPEED *
            input[i] * hidden_delta[j];
        }
}

/* Verify wether target, matches output */
int nnValidate() {
    int i;
    //printf ("Rounding: %lf - %lf\n",output[1], target[1]);
    for (i = 1; i < OUTPUT_SIZE; i++)
        if (round(output[i]) != round(target[i]))
            return FALSE;
    return TRUE;
}

/* Pretty print of output */
void nnOutput() {
    int i;
    for(i = 0; i < INPUT_SIZE; i++)
        printf("%lf, ", input[i]);
    printf("= %lf - %lf - ", output[1], target[1]);
    if (nnValidate() == TRUE)
        printf("OK");
    else
        printf("ERROR");
    printf("\n");
}

/* Pretty print of hidden knobs */
void nnHiddenOutput() {
    int i;
    for(i = 0; i < HIDDEN_SIZE; i++)
        printf("%lf, ", hidden[i]);
    printf(" - HIDDEN\n");
}


/* Pretty print of all weights */
void nnNeuronOutput() {
    int i,j;
    for (i = 0; i < INPUT_SIZE; i++)
        for(j = 0; j < HIDDEN_SIZE; j++)
            if (weight_ItoH[i][j] != WEIGHT_NOT_USED)
                printf("weight_ItoH[%i][%i] = %lf\n", i, j,
                weight_ItoH[i][j]);
    printf("---\n");
    for (i = 0; i < HIDDEN_SIZE; i++)
        for(j = 0; j < OUTPUT_SIZE; j++)
            if (weight_ItoH[i][j] != WEIGHT_NOT_USED)
                printf("weight_HtoO[%i][%i] = %lf\n", i, j,
                weight_HtoO[i][j]);
}

int nnReadInput(FILE * handle) {
    int i = 1;
    double finput;
    while (fscanf(handle, "%lf", &finput) != EOF) {
        if (i < INPUT_SIZE)
            input[i] = finput;
        else if (i < (INPUT_SIZE + OUTPUT_SIZE))
            target[i - INPUT_SIZE] = finput;

        /* Calc next input */
        i++;
        /* Skip hidden output knob */
        if (i == INPUT_SIZE)
            i++;
        if (i == (INPUT_SIZE + OUTPUT_SIZE))
            return TRUE;
    }

    /* Input not complete */
    return FALSE;
}

/* Verify quality of current network */
double nnQualityCheck(char * file) {
    double validate_total = 0;
    double validate_ok = 0;
    double validate_percent = 0;
    FILE * handle;

    handle = fopen(file,"r");
    while (nnReadInput(handle) == TRUE) {
            validate_total++;
            nnCalc();
            if (nnValidate() == TRUE)
                validate_ok++;
            //else
            //  nnOutput();
    }
    fclose(handle);
    validate_percent = (validate_ok / validate_total) * 100;
    printf("Validating: %.0lf/%.0lf - %.2lf %%\n",
    validate_ok,validate_total,validate_percent);

    return(validate_percent);
}

/* Main program */
int main (int argc, char * argv[]) {
    int i,training_total, training_best;
    double quality_max, quality;
    FILE * handle;

    /* Set the bias knob */
    input[0] = -1;
    hidden[0] = -1;
 
    /* Init set of all wights */
    //debugInit();
    fixedInit();
    //randInit();

    /* Set initial quality */
    quality_max = nnQualityCheck(file_quality);
    training_best = 0;
    training_total = 0;

    printf("Running neural network with following parameters\n");
    printf("Input  nodes    : %i\n", INPUT_SIZE);
    printf("Hidden nodes    : %i\n", HIDDEN_SIZE);
    printf("Output nodes    : %i\n", OUTPUT_SIZE);
    printf("Learning rate   : %lf\n", LEARN_SPEED);
    printf("Quality check   : %i\n", QUALITY_ROUND);
    printf("Initial quality : %lf %%\n", quality_max);
    /* Start training */
    //nnNeuronOutput();
    i = 1;
    handle = fopen(file_training,"r");
    while ( nnReadInput(handle) == TRUE) {
        training_total++;
        nnCalc();
        //nnOutput();
        //nnHiddenOutput();

        if (nnValidate() == FALSE) {
            nnTrain();
            //nnNeuronOutput();
        }

        /* Verifiy quality, stop training when quality is going down */
        if ((training_total % QUALITY_ROUND) == 0) {
            printf("Learned: %i - ", training_total);
            quality = nnQualityCheck(file_quality);
            if (quality > quality_max) {
                quality_max = quality;
                training_best = training_total;
            }
        }
    }
    fclose(handle);
    printf("Max quality: %.2lf%% at training round: %i\n", quality_max,
        training_best);
    quality = nnQualityCheck(file_validate);
    return(EX_OK);
}