From fb83fc30cf33cec1d155b3a63c338bbb64adb4e3 Mon Sep 17 00:00:00 2001
From: "Steinar H. Gunderson" <sgunderson@bigfoot.com>
Date: Sun, 17 Sep 2017 11:41:36 +0200
Subject: [PATCH] k-means instead of k-medoids; doesn't work as well, so just
 keep it here to be immediately reverted.

---
 qdc.cpp | 119 +++++++++++++++++++++++++++-----------------------------
 1 file changed, 57 insertions(+), 62 deletions(-)

diff --git a/qdc.cpp b/qdc.cpp
index b5594b7..c4e1ba4 100644
--- a/qdc.cpp
+++ b/qdc.cpp
@@ -374,80 +374,75 @@ double find_best_assignment(const int *medoids, int *assignment)
 	return current_score;
 }
 
-void find_optimal_stream_assignment(int base)
+double find_inv_sum(const SymbolStats &stats)
 {
-	double inv_sum[64];
-	for (unsigned i = 0; i < 64; ++i) {
-		double s = 0.0;
-		for (unsigned k = 0; k < 256; ++k) {
-			s += stats[i + base].freqs[k] + 0.5;
-		}
-		inv_sum[i] = 1.0 / s;
-	}
-
-	for (unsigned i = 0; i < 64; ++i) {
-		for (unsigned j = 0; j < 64; ++j) {
-			double d = 0.0;
-			for (unsigned k = 0; k < 256; ++k) {
-				double p1 = (stats[i + base].freqs[k] + 0.5) * inv_sum[i];
-				double p2 = (stats[j + base].freqs[k] + 0.5) * inv_sum[j];
-
-				// K-L divergence is asymmetric; this is a hack.
-				d += p1 * log(p1 / p2);
-				d += p2 * log(p2 / p1);
-			}
-			kl_dist[i][j] = d;
-			//printf("%.3f ", d);
-		}
-		//printf("\n");
+	double s = 0.0;
+	for (unsigned j = 0; j < NUM_SYMS; ++j) {
+		s += stats.freqs[j] + 0.5;
 	}
+	return 1.0 / s;
+}
 
-	// k-medoids init
-	int medoids[64];  // only the first NUM_CLUSTERS matter
-	bool is_medoid[64] = { false };
-	std::iota(medoids, medoids + 64, 0);
+void find_optimal_stream_assignment(int base)
+{
+	// k-means init; make random assignments
 	std::random_device rd;
 	std::mt19937 g(rd());
-	std::shuffle(medoids, medoids + 64, g);
-	for (int i = 0; i < NUM_CLUSTERS; ++i) {
-		printf("%d ", medoids[i]);
-		is_medoid[medoids[i]] = true;
+	std::uniform_int_distribution<> u(0, NUM_CLUSTERS - 1);
+	int assignment[64];
+	for (unsigned i = 0; i < 64; ++i) {
+		assignment[i] = u(g);
+	}	
+	double inv_sum_coeffs[64];
+	for (unsigned i = 0; i < 64; ++i) {
+		inv_sum_coeffs[i] = find_inv_sum(stats[i + base]);
 	}
-	printf("\n");
 
-	// assign each data point to the closest medoid
-	int assignment[64];
-	double current_score = find_best_assignment(medoids, assignment);
+	for (unsigned iter = 0; iter < 1000; ++iter) {
+		// make new clusters based on the current assignments
+		SymbolStats clusters[NUM_CLUSTERS];
+		for (unsigned i = 0; i < NUM_CLUSTERS; ++i) {
+			clusters[i].clear();
+		}
+		for (unsigned i = 0; i < 64; ++i) {
+			for (unsigned j = 0; j < NUM_SYMS; ++j) {
+				clusters[assignment[i]].freqs[j] += stats[i + base].freqs[j];
+			}
+		}
 
-	for (int i = 0; i < 1000; ++i) {
-		printf("iter %d\n", i);
+		double inv_sum_clusters[NUM_CLUSTERS];
+		for (unsigned i = 0; i < NUM_CLUSTERS; ++i) {
+			inv_sum_clusters[i] = find_inv_sum(clusters[i]);
+		}
+		
+		// find new assignments based on distance to the clusters
 		bool any_changed = false;
-		for (int m = 0; m < NUM_CLUSTERS; ++m) {
-			for (int o = 0; o < 64; ++o) {
-				if (is_medoid[o]) continue;
-				int old_medoid = medoids[m];
-				medoids[m] = o;
-
-				int new_assignment[64];
-				double candidate_score = find_best_assignment(medoids, new_assignment);
-
-				if (candidate_score < current_score) {
-					current_score = candidate_score;
-					memcpy(assignment, new_assignment, sizeof(assignment));
-
-					is_medoid[old_medoid] = false;
-					is_medoid[medoids[m]] = true;
-					printf("%f: ", current_score);
-					for (int i = 0; i < 64; ++i) {
-						printf("%d ", assignment[i]);
-					}
-					printf("\n");
-					any_changed = true;
-				} else {
-					medoids[m] = old_medoid;
+		double total_d = 0.0;
+		for (unsigned i = 0; i < 64; ++i) {
+			int best_assignment = -1;
+			double best_distance = HUGE_VAL;
+			for (unsigned j = 0; j < NUM_CLUSTERS; ++j) {
+				double d = 0.0;
+				for (unsigned k = 0; k < NUM_SYMS; ++k) {
+					double p1 = (clusters[j].freqs[k] + 0.5) * inv_sum_clusters[j];
+					double p2 = (stats[i + base].freqs[k] + 0.5) * inv_sum_coeffs[i];
+
+					// K-L divergence is asymmetric; this is a hack.
+					d += p1 * log(p1 / p2);
+					d += p2 * log(p2 / p1);
+				}
+				if (d < best_distance) {
+					best_assignment = j;
+					best_distance = d;
 				}
 			}
+			if (assignment[i] != best_assignment) {
+				any_changed = true;
+			}
+			assignment[i] = best_assignment;
+			total_d += best_distance;
 		}
+		printf("iter %u: %.3f\n", iter, total_d);
 		if (!any_changed) break;
 	}
 	printf("\n");
-- 
2.39.2