From cf7d665598e34e32d258775b583968e758c8f2fb Mon Sep 17 00:00:00 2001
From: tslil <tslil@posteo.de>
Date: Wed, 13 Jan 2021 23:41:32 -0500
Subject: Fixed minimax (!), fixed bugs in tak.c

With minimax of depth 1 the evaluation function seems alright with the
current method of training and generating weights
---
 src/ctaklm.c  |  2 +-
 src/pptdb.c   | 14 +++++---------
 src/train5.py | 22 ++++++++++++++++------
 3 files changed, 22 insertions(+), 16 deletions(-)

(limited to 'src')

diff --git a/src/ctaklm.c b/src/ctaklm.c
index 4060911..9d0f867 100644
--- a/src/ctaklm.c
+++ b/src/ctaklm.c
@@ -366,7 +366,7 @@ main(int argc, char **argv) {
 			int r = handle_turn(line);
 			if (r == EXIT_SUCCESS && won == 0xFF) {
 				fputs("Thinking [", stdout);
-				float minimax = ct1986_minimax(0, 2, 0);
+				float minimax = ct1986_minimax(0, 1, 0);
 				printf("] Result: %s (%.3f)\n",
 							 ct1986_ptn,
 							 minimax*100.0);
diff --git a/src/pptdb.c b/src/pptdb.c
index a67be1d..dbadfdf 100644
--- a/src/pptdb.c
+++ b/src/pptdb.c
@@ -4,10 +4,10 @@
 
 #include <tak.h>
 
-int  generate;
+float max_flats;
 uint64_t heights[16];
-float max_flats, outcome_black;
 FILE *training_fh = NULL;
+int generate, outcome_black;
 
 const int max_depth = 8;
 
@@ -107,11 +107,7 @@ parse_line(const char *pt, const ssize_t read) {
 		// Generate training data, not too early in the game
 		if (generate && ply + 5 > total_plies) {
 			write_input();
-			if (ply >= total_plies) {
-				if (outcome_black > 0.5) outcome_black = 1.0;
-				else outcome_black = 0.0;
-			}
-			fprintf(training_fh,"%.2f\n", outcome_black);
+			fprintf(training_fh,"%df\n", outcome_black);
 		}
 		// Parse next action
 		while (idx<read && pt[idx++]!=',');
@@ -164,8 +160,8 @@ 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 = 0.99;
-		else outcome_black = 0.01;
+		if (line[read-4] == '0') outcome_black = 1;
+		else outcome_black = 0;
 		// Parse the line
 		r = parse_line(line,read-4);
 		// Adjust counts if we're not generating training data
diff --git a/src/train5.py b/src/train5.py
index 71a8652..c3fec82 100644
--- a/src/train5.py
+++ b/src/train5.py
@@ -31,6 +31,17 @@ def truncated_pade_logistic(x):
                               divide(x, add(square(x), k(10.0))))))
     return minimum(k(1.0), maximum(k(0.0), divide(val, k(24.0))))
 
+
+# A cheaper version of tanh, x/6+25*x/(6*(2*x*x+5)),
+# clamped between -1 and 1.
+def truncated_pade_tanh(x):
+    val = add(divide(x, k(6.0)),
+              multiply(k(25.0),
+                       divide(x, multiply(k(6.0), add(k(5.0),
+                                                      multiply(k(2.0), square(x)))))))
+    return minimum(k(1.0), maximum(k(-1.0), val))
+
+
 def make_model(size, magic=[12, 9, 9]):
     # Our model for the stacks, a small CNN
     stack_shape = (size, size, 8)
@@ -59,7 +70,7 @@ def make_model(size, magic=[12, 9, 9]):
 
 def load_data(size):
     shape = (-1, size, size, 8)
-
+    # Load training data
     tr_fn = "training-"+str(size)+".csv"
     training_csv = pd.read_csv(tr_fn)  #.head(100000)
     training_data = training_csv
@@ -67,14 +78,13 @@ def load_data(size):
     training_flats_input = np.array(training_data.iloc[:, 0:2])
     training_input = [training_stack_input, training_flats_input]
     training_outcome = training_data.iloc[:, -1:]
-
+    # Load validation data
     val_fn = "validation-"+str(size)+".csv"
-    val_data = pd.read_csv(val_fn)  #.tail(20000)
+    val_data = pd.read_csv(val_fn).tail(20000)
     val_stack_input = np.array(val_data.iloc[:, 2:-1]).reshape(shape, order='F')
     val_flats_input = np.array(val_data.iloc[:, 0:2])
     val_input = [val_stack_input, val_flats_input]
     val_outcome = val_data.iloc[:, -1:]
-
     return [(training_input, training_outcome), (val_input, val_outcome)]
 
 def train(size, model, data, iterations=1, epochs=10):
@@ -150,8 +160,8 @@ def write_weights(model):
 
 
 data = load_data(5)
-model = make_model(5, [12, 9, 9])
-results, model = train(5, model, data, iterations=10, epochs=20)
+model = make_model(5, [12, 9, 10])
+results, model = train(5, model, data, iterations=1, epochs=10)
 print("\nScores")
 for i in range(len(results)):
     print("Iteration {0}: {1}".format(i+1, results[i]))
-- 
cgit v1.2.3