]> git.sesse.net Git - wloh/blobdiff - bayeswf.cpp
Make the Hessian calculation use the new all_matches vector.
[wloh] / bayeswf.cpp
index c54f72ab03d82cb7edc5c18f209e7620a3490ca1..487774ea394cb993c7b5c4499cd64f3fb6c30d55 100644 (file)
@@ -11,6 +11,7 @@
 using namespace std;
 
 #define PRIOR_MU 1500
+#define PRIOR_WEIGHT 1.0
 #define MAX_PLAYERS 4096
 #define DUMP_RAW 0
 
@@ -32,11 +33,12 @@ float prior_sigma = 70.0f;
  */
 
 struct match {
-       int other_player;
+       int player, other_player;
        int margin;
        float weight;
 };
 map<int, vector<match> > matches_for_player;
+vector<match> all_matches;
 
 void dump_scores(const vector<string> &players, const float *mu, const float *sigma, int num_players)
 {
@@ -74,8 +76,8 @@ void update_mu(float *mu, float *sigma, int player_num, const vector<match> &mat
        // Prior.
        {
                float inv_sigma2 = 1.0f / (prior_sigma * prior_sigma);
-               nom += PRIOR_MU * inv_sigma2;
-               denom += inv_sigma2;
+               nom += PRIOR_WEIGHT * PRIOR_MU * inv_sigma2;
+               denom += PRIOR_WEIGHT * inv_sigma2;
        }
 
        // All matches.
@@ -93,26 +95,19 @@ void update_mu(float *mu, float *sigma, int player_num, const vector<match> &mat
 
 void dump_raw(const float *mu, const float *sigma, int num_players)
 {
-       for (int i = 0; i < num_players; ++i) {
-               for (unsigned j = 0; j < matches_for_player[i].size(); ++j) {
-                       const match& m = matches_for_player[i][j];
-
-                       // Only count each match once.
-                       if (m.other_player <= i) {
-                               continue;
-                       }
-
-                       float mu1 = mu[i];
-                       float mu2 = mu[m.other_player];
-                       float sigma1 = sigma[i];
-                       float sigma2 = sigma[m.other_player];
-                       float sigma = sqrt(sigma1 * sigma1 + sigma2 * sigma2);
-                       float mu = mu1 - mu2;
-                       float x = m.margin;
-                       float w = m.weight;
-
-                       printf("%f %f\n", (x - mu) / sigma, w);
-               }
+       for (unsigned i = 0; i < all_matches.size(); ++i) {
+               const match& m = all_matches[i];
+
+               float mu1 = mu[m.player];
+               float mu2 = mu[m.other_player];
+               float sigma1 = sigma[m.player];
+               float sigma2 = sigma[m.other_player];
+               float sigma = sqrt(sigma1 * sigma1 + sigma2 * sigma2);
+               float mu = mu1 - mu2;
+               float x = m.margin;
+               float w = m.weight;
+
+               printf("%f %f\n", (x - mu) / sigma, w);
        }
 }
 
@@ -158,24 +153,17 @@ void update_sigma(float *mu, float *sigma, int player_num, const vector<match> &
 void update_global_sigma(float *mu, float *sigma, int num_players)
 {
        float nom = 0.0f, denom = 0.0f;
-       for (int i = 0; i < num_players; ++i) {
-               for (unsigned j = 0; j < matches_for_player[i].size(); ++j) {
-                       const match& m = matches_for_player[i][j];
-
-                       // Only count each match once.
-                       if (m.other_player <= i) {
-                               continue;
-                       }
-
-                       float mu1 = mu[i];
-                       float mu2 = mu[m.other_player];
-                       float mu = mu1 - mu2;
-                       float x = m.margin;
-                       float w = m.weight;
-
-                       nom += w * ((x - mu) * (x - mu));
-                       denom += w;
-               }
+       for (unsigned i = 0; i < all_matches.size(); ++i) {
+               const match& m = all_matches[i];
+
+               float mu1 = mu[m.player];
+               float mu2 = mu[m.other_player];
+               float mu = mu1 - mu2;
+               float x = m.margin;
+               float w = m.weight;
+
+               nom += w * ((x - mu) * (x - mu));
+               denom += w;
        }
 
        float best_sigma = sqrt(nom / denom) / sqrt(2.0f);  // Divide evenly between the two players.
@@ -195,23 +183,49 @@ void update_prior_sigma(float *mu, float *sigma, int num_players)
 {
        float nom = 0.0f, denom = 0.0f;
        for (int i = 0; i < num_players; ++i) {
-               for (unsigned j = 0; j < matches_for_player[i].size(); ++j) {
-                       const match& m = matches_for_player[i][j];
+               float mu1 = mu[i];
 
-                       // Only count each match once.
-                       if (m.other_player <= i) {
-                               continue;
-                       }
+               nom += ((mu1 - PRIOR_MU) * (mu1 - PRIOR_MU));
+               denom += 1.0f;
+       }
 
-                       float mu1 = mu[i];
+       prior_sigma = sqrt(nom / denom);
+       if (!(prior_sigma > 40.0f)) {
+               prior_sigma = 40.0f;
+       }
+}
 
-                       float w = m.weight;
-                       nom += w * ((mu1 - PRIOR_MU) * (mu1 - PRIOR_MU));
-                       denom += w * 1.0f;
-               }
+float compute_logl(float z)
+{
+       return -0.5 * (log(2.0f / M_PI) + z * z);
+}
+
+float compute_total_logl(float *mu, float *sigma, int num_players)
+{
+       float total_logl = 0.0f;
+
+       // Prior.
+       for (int i = 0; i < num_players; ++i) {
+               total_logl += PRIOR_WEIGHT * compute_logl((mu[i] - PRIOR_MU) / prior_sigma);
        }
 
-       prior_sigma = sqrt(nom / denom);
+       // Matches.
+       for (unsigned i = 0; i < all_matches.size(); ++i) {
+               const match& m = all_matches[i];
+
+               float mu1 = mu[m.player];
+               float mu2 = mu[m.other_player];
+               float sigma1 = sigma[m.player];
+               float sigma2 = sigma[m.other_player];
+               float sigma = sqrt(sigma1 * sigma1 + sigma2 * sigma2);
+               float mu = mu1 - mu2;
+               float x = m.margin;
+               float w = m.weight;
+
+               total_logl += w * compute_logl((x - mu) / sigma);
+       }
+
+       return total_logl;
 }
 
 /*
@@ -227,20 +241,23 @@ void construct_hessian(const float *mu, const float *sigma, int num_players)
 {
        memset(hessian, 0, sizeof(hessian));
 
-       for (int i = 0; i < num_players; ++i) {
-               double sigma1 = sigma[i];
+       for (unsigned i = 0; i < all_matches.size(); ++i) {
+               const match &m = all_matches[i];
 
-               for (unsigned k = 0; k < matches_for_player[i].size(); ++k) {
-                       int j = matches_for_player[i][k].other_player;
+               int p1 = m.player;
+               int p2 = m.other_player;
 
-                       double sigma2 = sigma[j];
-                       double sigma_sq = sigma1 * sigma1 + sigma2 * sigma2;
+               double sigma1 = sigma[m.player];
+               double sigma2 = sigma[m.other_player];
 
-                       float w = matches_for_player[i][k].weight;
+               double sigma_sq = sigma1 * sigma1 + sigma2 * sigma2;
+               float w = m.weight;
 
-                       hessian[i][j] -= w / sigma_sq;
-                       hessian[i][i] += w / sigma_sq;
-               }
+               hessian[p1][p2] -= w / sigma_sq;
+               hessian[p2][p1] -= w / sigma_sq;
+
+               hessian[p1][p1] += w / sigma_sq;
+               hessian[p2][p2] += w / sigma_sq;
        }
 
        for (int i = 0; i < num_players; ++i) {
@@ -301,23 +318,27 @@ int main(int argc, char **argv)
                }
 
                match m1;
+               m1.player = player_map[pl1];
                m1.other_player = player_map[pl2];
                m1.margin = score1 - score2;
                m1.weight = weight;
                matches_for_player[player_map[pl1]].push_back(m1);
 
                match m2;
+               m2.player = player_map[pl2];
                m2.other_player = player_map[pl1];
                m2.margin = score2 - score1;
                m2.weight = weight;
                matches_for_player[player_map[pl2]].push_back(m2);
+
+               all_matches.push_back(m1);
        }
 
        float mu[MAX_PLAYERS];
        float sigma[MAX_PLAYERS];
 
        for (int i = 0; i < num_players; ++i) {
-               mu[i] = 1500.0f;
+               mu[i] = PRIOR_MU;
                sigma[i] = 70.0f / sqrt(2.0f);
        }
 
@@ -358,6 +379,9 @@ int main(int argc, char **argv)
        printf("%f -2\n", sigma[0]);
        printf("%f -3\n", prior_sigma);
 
+       float total_logl = compute_total_logl(mu, sigma, num_players);
+       printf("%f -4\n", total_logl);
+
 //     construct_hessian(mu, sigma, num_players);
 #endif
 }