14 #define PRIOR_WEIGHT 1.0
15 #define MAX_PLAYERS 4096
18 float mu[MAX_PLAYERS];
19 float sigma[MAX_PLAYERS];
20 float prior_sigma = 70.0f;
25 * L(mu_vec, sigma_vec, matches) = product[ L(mu_A, sigma_A, mu_B, sigma_B, score_AB - score_BA) ]
26 * log-likelihood = sum[ log( L(mu_A, sigma_A, mu_B, sigma_B, score_AB - score_BA) ) ]
28 * L(mu1, sigma1, mu2, sigma2, score2 - score1) = sigmoid(mu2 - mu1, sqrt(sigma1² + sigma2²), (score2 - score1))
30 * pdf := 1/(sigma * sqrt(2*Pi)) * exp(-(x - mu)^2 / (2 * sigma^2));
31 * pdfs := subs({ mu = mu1 - mu2, sigma = sqrt(sigma1^2 + sigma2^2) }, pdf);
32 * diff(log(pdfs), mu1);
36 int player, other_player;
40 map<int, vector<match> > matches_for_player;
41 vector<match> all_matches;
43 void dump_scores(const vector<string> &players, const float *mu, const float *sigma, int num_players)
46 for (int i = 0; i < num_players; ++i) {
47 printf("%s=[%5.1f, %4.1f] ", players[i].c_str(), mu[i], sigma[i]);
51 for (int i = 0; i < num_players; ++i) {
52 printf("%5.1f ", mu[i]);
56 for (int i = 0; i < num_players; ++i) {
57 printf("%f %s\n", mu[i], players[i].c_str());
63 * diff(logL, mu1) = -w * (mu1 - mu2 - x) / sigma_c^2
64 * maximizer for mu1 is given by: sum_i[ (w_i/sigma_c_i)^2 (mu1 - mu2_i - x_i) ] = 0
65 * sum_i[ (w_i/sigma_c_i)^2 mu1 ] = sum_i [ (w_i/sigma_c_i)^2 ( mu2_i + x_i ) ]
66 * mu1 = sum_i [ (w_i/sigma_c_i)^2 ( mu2_i + x_i ) ] / sum_i[ (w_i/sigma_c_i)^2 ]
68 void update_mu(float *mu, float *sigma, int player_num, const vector<match> &matches)
70 if (matches.empty()) {
74 float nom = 0.0f, denom = 0.0f;
78 float inv_sigma2 = 1.0f / (prior_sigma * prior_sigma);
79 nom += PRIOR_WEIGHT * PRIOR_MU * inv_sigma2;
80 denom += PRIOR_WEIGHT * inv_sigma2;
84 for (unsigned i = 0; i < matches.size(); ++i) {
85 float sigma1 = sigma[player_num];
86 float sigma2 = sigma[matches[i].other_player];
87 float inv_sigma_c2 = matches[i].weight / (sigma1 * sigma1 + sigma2 * sigma2);
88 float x = matches[i].margin; // / 70.0f;
90 nom += (mu[matches[i].other_player] + x) * inv_sigma_c2;
91 denom += inv_sigma_c2;
93 mu[player_num] = nom / denom;
96 void dump_raw(const float *mu, const float *sigma, int num_players)
98 for (unsigned i = 0; i < all_matches.size(); ++i) {
99 const match& m = all_matches[i];
101 float mu1 = mu[m.player];
102 float mu2 = mu[m.other_player];
103 float sigma1 = sigma[m.player];
104 float sigma2 = sigma[m.other_player];
105 float sigma = sqrt(sigma1 * sigma1 + sigma2 * sigma2);
106 float mu = mu1 - mu2;
110 printf("%f %f\n", (x - mu) / sigma, w);
115 * diff(logL, sigma1) = sigma1 (-sigma1² - sigma2² + (x - mu)²) / sigma_c²
116 * maximizer for sigma1 is given by: sum_i[ (1/sigma_c_i)² sigma1 ((x - mu)² - (sigma1² + sigma2²) ] = 0
117 * sum_i[ (x - mu)² - sigma1² - sigma2² ] = 0 |: sigma1 != 0, sigma2 != 0
118 * sum_i[ (x - mu)² - sigma2² ] = sum[ sigma1² ]
119 * sigma1 = sqrt( sum_i[ (x - mu)² - sigma2² ] / N )
121 void update_sigma(float *mu, float *sigma, int player_num, const vector<match> &matches)
123 if (matches.size() < 2) {
128 for (unsigned i = 0; i < matches.size(); ++i) {
129 float mu1 = mu[player_num];
130 float mu2 = mu[matches[i].other_player];
131 float mu = mu1 - mu2;
132 float sigma2 = sigma[matches[i].other_player];
133 float x = matches[i].margin;
135 //fprintf(stderr, "x=%f mu=%f sigma2=%f add %f-%f = %f\n", x, mu, sigma2, (x-mu)*(x-mu), sigma2*sigma2, (x - mu) * (x - mu) - sigma2 * sigma2);
136 sum += (x - mu) * (x - mu) - sigma2 * sigma2;
142 //fprintf(stderr, "sum=%f\n", sum);
143 sigma[player_num] = sqrt(sum / matches.size());
147 * diff(logL, sigma) = w ( (x - mu)² - sigma² ) / sigma³
148 * maximizer for sigma is given by: sum_i[ (w_i/sigma)³ ((x - mu)² - sigma²) ] = 0
149 * sum_i[ w_i ( (x - mu)² - sigma² ) ] = 0 |: sigma != 0
150 * sum_i[ w_i (x - mu)² ] = sum[ w_i sigma² ]
151 * sigma = sqrt( sum_i[ w_i (x - mu)² ] / sum[w_i] )
153 void update_global_sigma(float *mu, float *sigma, int num_players)
155 float nom = 0.0f, denom = 0.0f;
156 for (unsigned i = 0; i < all_matches.size(); ++i) {
157 const match& m = all_matches[i];
159 float mu1 = mu[m.player];
160 float mu2 = mu[m.other_player];
161 float mu = mu1 - mu2;
165 nom += w * ((x - mu) * (x - mu));
169 float best_sigma = sqrt(nom / denom) / sqrt(2.0f); // Divide evenly between the two players.
170 for (int i = 0; i < num_players; ++i) {
171 sigma[i] = best_sigma;
176 * diff(priorlogL, sigma) = w ( (x - mu)² - sigma² ) / sigma³
177 * maximizer for sigma is given by: sum_i[ (w_i/sigma)³ ((x - mu)² - sigma²) ] = 0
178 * sum_i[ w_i ( (x - mu)² - sigma² ) ] = 0 |: sigma != 0
179 * sum_i[ w_i (x - mu)² ] = sum[ w_i sigma² ]
180 * sigma = sqrt( sum_i[ w_i (x - mu)² ] / sum[w_i] )
182 void update_prior_sigma(float *mu, float *sigma, int num_players)
184 float nom = 0.0f, denom = 0.0f;
185 for (int i = 0; i < num_players; ++i) {
188 nom += ((mu1 - PRIOR_MU) * (mu1 - PRIOR_MU));
192 prior_sigma = sqrt(nom / denom);
193 if (!(prior_sigma > 40.0f)) {
198 float compute_logl(float z)
200 return -0.5 * (log(2.0f / M_PI) + z * z);
203 float compute_total_logl(float *mu, float *sigma, int num_players)
205 float total_logl = 0.0f;
208 for (int i = 0; i < num_players; ++i) {
209 total_logl += PRIOR_WEIGHT * compute_logl((mu[i] - PRIOR_MU) / prior_sigma);
213 for (unsigned i = 0; i < all_matches.size(); ++i) {
214 const match& m = all_matches[i];
216 float mu1 = mu[m.player];
217 float mu2 = mu[m.other_player];
218 float sigma1 = sigma[m.player];
219 float sigma2 = sigma[m.other_player];
220 float sigma = sqrt(sigma1 * sigma1 + sigma2 * sigma2);
221 float mu = mu1 - mu2;
225 total_logl += w * compute_logl((x - mu) / sigma);
232 * Compute Hessian matrix of the negative log-likelihood, ie. for each term in logL:
234 * M_ij = D_i D_j (- logL) = -w / sigma² for i != j
235 * w / sigma² for i == j
237 * Note that this does not depend on mu or the margin at all.
239 double hessian[MAX_PLAYERS][MAX_PLAYERS];
240 void construct_hessian(const float *mu, const float *sigma, int num_players)
242 memset(hessian, 0, sizeof(hessian));
244 for (unsigned i = 0; i < all_matches.size(); ++i) {
245 const match &m = all_matches[i];
248 int p2 = m.other_player;
250 double sigma1 = sigma[m.player];
251 double sigma2 = sigma[m.other_player];
253 double sigma_sq = sigma1 * sigma1 + sigma2 * sigma2;
256 hessian[p1][p2] -= w / sigma_sq;
257 hessian[p2][p1] -= w / sigma_sq;
259 hessian[p1][p1] += w / sigma_sq;
260 hessian[p2][p2] += w / sigma_sq;
263 for (int i = 0; i < num_players; ++i) {
264 for (int j = 0; j < num_players; ++j) {
265 printf("%.12f ", hessian[i][j]);
271 int main(int argc, char **argv)
274 if (scanf("%d", &num_players) != 1) {
275 fprintf(stderr, "Could't read number of players\n");
279 if (num_players > MAX_PLAYERS) {
280 fprintf(stderr, "Max %d players supported\n", MAX_PLAYERS);
284 vector<string> players;
285 map<string, int> player_map;
287 for (int i = 0; i < num_players; ++i) {
289 if (scanf("%s", buf) != 1) {
290 fprintf(stderr, "Couldn't read player %d\n", i);
294 players.push_back(buf);
300 char pl1[256], pl2[256];
304 if (scanf("%s %s %d %d %f", pl1, pl2, &score1, &score2, &weight) != 5) {
305 //fprintf(stderr, "Read %d matches.\n", num_matches);
311 if (player_map.count(pl1) == 0) {
312 fprintf(stderr, "Unknown player '%s'\n", pl1);
315 if (player_map.count(pl2) == 0) {
316 fprintf(stderr, "Unknown player '%s'\n", pl2);
321 m1.player = player_map[pl1];
322 m1.other_player = player_map[pl2];
323 m1.margin = score1 - score2;
325 matches_for_player[player_map[pl1]].push_back(m1);
328 m2.player = player_map[pl2];
329 m2.other_player = player_map[pl1];
330 m2.margin = score2 - score1;
332 matches_for_player[player_map[pl2]].push_back(m2);
334 all_matches.push_back(m1);
337 float mu[MAX_PLAYERS];
338 float sigma[MAX_PLAYERS];
340 for (int i = 0; i < num_players; ++i) {
342 sigma[i] = 70.0f / sqrt(2.0f);
345 for (int j = 0; j < 1000; ++j) {
346 float old_mu[MAX_PLAYERS];
347 float old_sigma[MAX_PLAYERS];
348 float old_prior_sigma = prior_sigma;
349 memcpy(old_mu, mu, sizeof(mu));
350 memcpy(old_sigma, sigma, sizeof(sigma));
351 for (int i = 0; i < num_players; ++i) {
352 update_mu(mu, sigma, i, matches_for_player[i]);
354 update_global_sigma(mu, sigma, num_players);
355 update_prior_sigma(mu, sigma, num_players);
356 /* for (int i = 0; i < num_players; ++i) {
357 update_sigma(mu, sigma, i, matches_for_player[i]);
358 dump_scores(players, mu, sigma, num_players);
361 float sumdiff = 0.0f;
362 for (int i = 0; i < num_players; ++i) {
363 sumdiff += (mu[i] - old_mu[i]) * (mu[i] - old_mu[i]);
364 sumdiff += (sigma[i] - old_sigma[i]) * (sigma[i] - old_sigma[i]);
366 sumdiff += (prior_sigma - old_prior_sigma) * (prior_sigma - old_prior_sigma);
367 if (sumdiff < EPSILON) {
368 //fprintf(stderr, "Converged after %d iterations. Stopping.\n", j);
369 printf("%d -1\n", j + 1);
375 dump_raw(mu, sigma, num_players);
377 dump_scores(players, mu, sigma, num_players);
378 //fprintf(stderr, "Optimal sigma: %f (two-player: %f)\n", sigma[0], sigma[0] * sqrt(2.0f));
379 printf("%f -2\n", sigma[0]);
380 printf("%f -3\n", prior_sigma);
382 float total_logl = compute_total_logl(mu, sigma, num_players);
383 printf("%f -4\n", total_logl);
385 // construct_hessian(mu, sigma, num_players);