Change the training data generation a little

Although it pains me to say it, ``label smoothing'' appears to be
actually work. I'm also currently experimenting with training simply
against _all_ games, instead of only bot matches. Once the training
finishes i'll pit cttei against itself with old and new weights,
hopefully there'll be a noticeable improvement.

1 files changed, 8 insertions, 10 deletions

diff --git a/src/pptdb.c b/src/pptdb.c
index bb94987..286678d 100644
--- a/src/pptdb.c
+++ b/src/pptdb.c
@@ -24,10 +24,10 @@
 
 #include <tak.h>
 
-float max_flats;
+int generate;
 uint64_t heights[16];
 FILE *training_fh = NULL;
-int generate, outcome_black;
+float max_flats, outcome_black;
 
 const int max_depth = 6;
 
@@ -71,7 +71,7 @@ write_input(const int dx, const int dy, const uint8_t swap) {
 			}
 		}
 	}
-	fprintf(training_fh,"%d\n", outcome_black);
+	fprintf(training_fh,"%.1f\n", outcome_black);
 }
 
 // Warning: performs _no_ checks on input whatsoever
@@ -136,11 +136,9 @@ parse_line(const char *pt, const ssize_t read) {
 		}
 		// Generate training data, not too early in the game, all
 		// orientations
-		if (generate && ply + 4 >= total_plies) {
-			write_input(-1, -1, 0); write_input(-1, -1, 1);
-			write_input(-1, +1, 0);	write_input(-1, +1, 1);
-			write_input(+1, -1, 0);	write_input(+1, -1, 1);
-			write_input(+1, +1, 0);	write_input(+1, +1, 1);
+		if (generate && ply + 2 >= total_plies) {
+#define RANDPM1 ((rand()&1)?-1:+1)
+			write_input(RANDPM1, RANDPM1, rand()&1);
 		}
 		// Parse next action
 		while (idx<read && pt[idx++]!=',');
@@ -198,8 +196,8 @@ int main(int argc, char **argv) {
 		// Reset everything
 		reset_state(size);
 		// Store the outcome of this game. Black win = 1
-		if (line[read-4] == '0') outcome_black = 1;
-		else outcome_black = -1;
+		if (line[read-4] == '0') outcome_black = 0.9;
+		else outcome_black = -0.9;
 		// Parse the line
 		r = parse_line(line,read-4);
 		// Adjust counts if we're not generating training data