close Warning: Can't use blame annotator: 
No changeset 422 in the repository

source: liacs/ai/poker/nn.c@ 422

Last change on this file since 422 was 2, checked in by Rick van der Zwet, 15 years ago

Initial import of data of old repository ('data') worth keeping (e.g. tracking
means of URL access statistics)
File size: 8.6 KB
RevLine 
1/*
2 * Rick van der Zwet
3 * 0433373
4 * OS Assigment 3
5 * Licence: BSD
6 * $Id: nn.c 557 2008-04-08 22:57:09Z rick $
7*/
8
9#include <sysexits.h>
10#include <stdio.h>
11#include <stdlib.h>
12#include <math.h>
13#include <time.h>
14
15/* NOTE: All first knobs are bias knobs or hidden stale knobs
16 * - Validation is done using rounding, please make outputs discrete or
17 * alter validation function
18 */
19
20/* Allow uniform and easy calls at functions */
21#define TRUE 1
22#define FALSE 0
23
24
25/* Network variables */
26/*NOTE: first node is 'hidden' bias knob */
27#ifndef INPUT_SIZE
28#define INPUT_SIZE 11
29#endif
30
31/*NOTE: first node is 'hidden' bias knob */
32#ifndef HIDDEN_SIZE
33#define HIDDEN_SIZE 11
34#endif
35
36/*NOTE: first node is 'hidden' 'lame' knob */
37#ifndef OUTPUT_SIZE
38#define OUTPUT_SIZE 11
39#endif
40
41/* Learn speed alpha of network */
42#ifndef LEARN_SPEED
43#define LEARN_SPEED 0.5
44#endif
45
46/* After QUALITY_ROUND trainingset check quality of network */
47#define QUALITY_ROUND 100
48
49/* Training set, to be used to train network */
50char * file_training = "data/training.txt";
51/* Validation set, to be used to test end result of network */
52char * file_validate = "data/validate.txt";
53/* Quality set, to be used to do quick testing whether network is
54 * improving
55 */
56char * file_quality = "data/quality.txt";
57
58/* Globally defined arrays, which represent the network */
59double hidden[HIDDEN_SIZE];
60double input[INPUT_SIZE];
61double output[OUTPUT_SIZE];
62double target[OUTPUT_SIZE];
63double weight_HtoO[HIDDEN_SIZE][OUTPUT_SIZE];
64double weight_ItoH[INPUT_SIZE][HIDDEN_SIZE];
65
66#define WEIGHT_NOT_USED -99999
67
68void stdInit() {
69 int i;
70 /* Should never change, been using */
71 for (i = 0; i < INPUT_SIZE; i++)
72 weight_ItoH[i][0] = WEIGHT_NOT_USED;
73 for (i = 0; i < HIDDEN_SIZE; i++)
74 weight_HtoO[i][0] = WEIGHT_NOT_USED;
75}
76
77
78/* Random init of weights */
79void randInit() {
80 int i,j;
81
82 /* Different numbers every call */
83 srandom(time(NULL));
84
85 for (i = 0; i < INPUT_SIZE; i++)
86 for ( j = 1; j < HIDDEN_SIZE; j++) {
87 weight_ItoH[i][j] = (double)(random() % 100) / 100;
88 }
89
90 for (i = 0; i < HIDDEN_SIZE; i++)
91 for (j = 1; j < OUTPUT_SIZE; j++)
92 weight_HtoO[i][j] = (double)(random() % 100) / 100;
93
94 stdInit();
95}
96
97/* Fixed init of weights */
98void fixedInit() {
99 int i,j;
100 for (i = 0; i < INPUT_SIZE; i++)
101 for ( j = 1; j < HIDDEN_SIZE; j++) {
102 weight_ItoH[i][j] = 0.5;
103 }
104
105 for (i = 0; i < HIDDEN_SIZE; i++)
106 for (j = 1; j < OUTPUT_SIZE; j++)
107 weight_HtoO[i][j] = 0.5;
108
109 stdInit();
110}
111
112/* Define exact wights, used for debugging calculations
113 * other flags INPUT = 2, HIDDEN = 2, OUTPUT = 1
114 */
115void debugInit() {
116 stdInit();
117 weight_ItoH[0][1] = 1;
118 weight_ItoH[0][2] = 1;
119 weight_ItoH[1][1] = 0.62;
120 weight_ItoH[1][2] = 0.42;
121 weight_ItoH[2][1] = 0.55;
122 weight_ItoH[2][2] = -0.17;
123
124 weight_HtoO[0][1] = 1;
125 weight_HtoO[1][1] = 0.35;
126 weight_HtoO[2][1] = 0.81;
127}
128
129/* calculate Aj's and Ai's (outputs) */
130void nnCalc() {
131 int i,j;
132 double total;
133 for (i = 1; i < HIDDEN_SIZE; i++) {
134 total = 0;
135 for (j = 0; j < INPUT_SIZE; j++)
136 total += weight_ItoH[j][i] * input[j];
137 hidden[i] = 1 / ( 1 + exp(total * (-1)));
138 }
139
140 for (i = 1; i < OUTPUT_SIZE; i++) {
141 total = 0;
142 for (j = 0; j < HIDDEN_SIZE; j++)
143 total += weight_HtoO[j][i] * hidden[j];
144 output[i] = 1 / ( 1 + exp(total * (-1)));
145 }
146
147}
148
149/* train network, NOTE: nnCalc needs to be called first */
150void nnTrain() {
151 int i,j;
152 double hidden_delta[HIDDEN_SIZE];
153 double output_delta[OUTPUT_SIZE];
154 double output_error[OUTPUT_SIZE];
155 double hidden_sum_delta[HIDDEN_SIZE];
156
157 for (i = 1; i < OUTPUT_SIZE; i++) {
158 output_error[i] = target[i] - output[i];
159 output_delta[i] = output_error[i] * output[i] * (1 - output[i]);
160 }
161
162 for (i = 0; i < HIDDEN_SIZE; i++) {
163 hidden_sum_delta[i] = 0;
164 for (j = 1; j < OUTPUT_SIZE; j++)
165 hidden_sum_delta[i] += weight_HtoO[i][j] * output_delta[j];
166 hidden_delta[i] = hidden[i] * (1 - hidden[i]) *
167 hidden_sum_delta[i];
168 }
169
170 for (i = 0; i < HIDDEN_SIZE; i++)
171 for (j = 1; j < OUTPUT_SIZE; j++) {
172 weight_HtoO[i][j] = weight_HtoO[i][j] + LEARN_SPEED *
173 hidden[i] * output_delta[j];
174 }
175
176 for (i = 0; i < INPUT_SIZE; i++)
177 for (j = 1; j < HIDDEN_SIZE; j++) {
178 weight_ItoH[i][j] = weight_ItoH[i][j] + LEARN_SPEED *
179 input[i] * hidden_delta[j];
180 }
181}
182
183/* Verify wether target, matches output */
184int nnValidate() {
185 int i;
186 //printf ("Rounding: %lf - %lf\n",output[1], target[1]);
187 for (i = 1; i < OUTPUT_SIZE; i++)
188 if (round(output[i]) != round(target[i]))
189 return FALSE;
190 return TRUE;
191}
192
193/* Pretty print of output */
194void nnOutput() {
195 int i;
196 for(i = 0; i < INPUT_SIZE; i++)
197 printf("%lf, ", input[i]);
198 printf("= %lf - %lf - ", output[1], target[1]);
199 if (nnValidate() == TRUE)
200 printf("OK");
201 else
202 printf("ERROR");
203 printf("\n");
204}
205
206/* Pretty print of hidden knobs */
207void nnHiddenOutput() {
208 int i;
209 for(i = 0; i < HIDDEN_SIZE; i++)
210 printf("%lf, ", hidden[i]);
211 printf(" - HIDDEN\n");
212}
213
214
215/* Pretty print of all weights */
216void nnNeuronOutput() {
217 int i,j;
218 for (i = 0; i < INPUT_SIZE; i++)
219 for(j = 0; j < HIDDEN_SIZE; j++)
220 if (weight_ItoH[i][j] != WEIGHT_NOT_USED)
221 printf("weight_ItoH[%i][%i] = %lf\n", i, j,
222 weight_ItoH[i][j]);
223 printf("---\n");
224 for (i = 0; i < HIDDEN_SIZE; i++)
225 for(j = 0; j < OUTPUT_SIZE; j++)
226 if (weight_ItoH[i][j] != WEIGHT_NOT_USED)
227 printf("weight_HtoO[%i][%i] = %lf\n", i, j,
228 weight_HtoO[i][j]);
229}
230
231int nnReadInput(FILE * handle) {
232 int i = 1;
233 double finput;
234 while (fscanf(handle, "%lf", &finput) != EOF) {
235 if (i < INPUT_SIZE)
236 input[i] = finput;
237 else if (i < (INPUT_SIZE + OUTPUT_SIZE))
238 target[i - INPUT_SIZE] = finput;
239
240 /* Calc next input */
241 i++;
242 /* Skip hidden output knob */
243 if (i == INPUT_SIZE)
244 i++;
245 if (i == (INPUT_SIZE + OUTPUT_SIZE))
246 return TRUE;
247 }
248
249 /* Input not complete */
250 return FALSE;
251}
252
253/* Verify quality of current network */
254double nnQualityCheck(char * file) {
255 double validate_total = 0;
256 double validate_ok = 0;
257 double validate_percent = 0;
258 FILE * handle;
259
260 handle = fopen(file,"r");
261 while (nnReadInput(handle) == TRUE) {
262 validate_total++;
263 nnCalc();
264 if (nnValidate() == TRUE)
265 validate_ok++;
266 //else
267 // nnOutput();
268 }
269 fclose(handle);
270 validate_percent = (validate_ok / validate_total) * 100;
271 printf("Validating: %.0lf/%.0lf - %.2lf %%\n",
272 validate_ok,validate_total,validate_percent);
273
274 return(validate_percent);
275}
276
277/* Main program */
278int main (int argc, char * argv[]) {
279 int i,training_total, training_best;
280 double quality_max, quality;
281 FILE * handle;
282
283 /* Set the bias knob */
284 input[0] = -1;
285 hidden[0] = -1;
286
287 /* Init set of all wights */
288 //debugInit();
289 fixedInit();
290 //randInit();
291
292 /* Set initial quality */
293 quality_max = nnQualityCheck(file_quality);
294 training_best = 0;
295 training_total = 0;
296
297 printf("Running neural network with following parameters\n");
298 printf("Input nodes : %i\n", INPUT_SIZE);
299 printf("Hidden nodes : %i\n", HIDDEN_SIZE);
300 printf("Output nodes : %i\n", OUTPUT_SIZE);
301 printf("Learning rate : %lf\n", LEARN_SPEED);
302 printf("Quality check : %i\n", QUALITY_ROUND);
303 printf("Initial quality : %lf %%\n", quality_max);
304 /* Start training */
305 //nnNeuronOutput();
306 i = 1;
307 handle = fopen(file_training,"r");
308 while ( nnReadInput(handle) == TRUE) {
309 training_total++;
310 nnCalc();
311 //nnOutput();
312 //nnHiddenOutput();
313
314 if (nnValidate() == FALSE) {
315 nnTrain();
316 //nnNeuronOutput();
317 }
318
319 /* Verifiy quality, stop training when quality is going down */
320 if ((training_total % QUALITY_ROUND) == 0) {
321 printf("Learned: %i - ", training_total);
322 quality = nnQualityCheck(file_quality);
323 if (quality > quality_max) {
324 quality_max = quality;
325 training_best = training_total;
326 }
327 }
328 }
329 fclose(handle);
330 printf("Max quality: %.2lf%% at training round: %i\n", quality_max,
331 training_best);
332 quality = nnQualityCheck(file_validate);
333 return(EX_OK);
334}
335
Note: See TracBrowser for help on using the repository browser.