From 241156da6f35cb888253e6b8814e7b61a76b4fc7 Mon Sep 17 00:00:00 2001
From: tslil <tslil@posteo.de>
Date: Sun, 30 Aug 2026 19:46:07 +0100
Subject: TD(lambda), and a big rework

NOTE: 6x6 weights are as-yet untrained

On the performance side:
* switch from linked-list chaining hash table to direct mapping
* cache check_win by hash
* persist transposition tables across plies (this was a big deal and difficult to puzzle out)
* add align directives to the network floats
* prepare_nn_input gained a small LUT to avoid branching (i suppose you might call this 1D embedding)
* apply mtune=native for the native builds

On the correctness side:
* Zobrist has now includes the ply parity
* the network now predicts something in [-1, 1] using tanh (approximated by polynomials)
* the training object is now MSE because we're learning a value estimate
* use .clangd instead of compile_commands.json

On the training side:
* completely forgo human-derived data set
* implement td(lambda), self-play, tournaments
---
 include/nn.h | 83 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 78 insertions(+), 5 deletions(-)

(limited to 'include/nn.h')

diff --git a/include/nn.h b/include/nn.h
index 5d1bf15..18e222a 100644
--- a/include/nn.h
+++ b/include/nn.h
@@ -1,5 +1,5 @@
 /*
-  This file is part of ct.
+  This file is part of ctak.
 
   This program is free software: you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
@@ -12,12 +12,85 @@
   General Public License for more details.
 
   You should have received a copy of the GNU General Public License
-  along with ct. If not, see <https://www.gnu.org/licenses/>.
+  along with ctak. If not, see <https://www.gnu.org/licenses/>.
 */
 
+#ifndef NN_H
+#define NN_H
+
 #include <tak.h>
+#include "weights_5.h"
+#include "weights_6.h"
+
+// ===================================================================
+// Helpers
+// ===================================================================
+
+void nn_prepare_input(float *cur_board, tak_state_p state);
+
+#define RELU(x) ((x) = ((x) < 0) ? 0 : (x))
+
+/*
+ * tanh via a degree-11 odd polynomial, least-squares fit of tanh on [-3, 3]
+ * (max absolute error ~4e-3), saturating to +-1 outside that interval.
+ */
+static inline float fast_tanh(const float x) {
+  if (x > 3.0f)
+    return 1.0f;
+  if (x < -3.0f)
+    return -1.0f;
+  const float x2 = x * x;
+  return x * (0.997136032f + x2 * (-0.305215124f + x2 *
+                                   (0.0832466371f + x2 *
+                                    (-0.0139947924f + x2 *
+                                     (0.00123106314f + x2 * -4.29748644e-05f)))));
+}
+
+// ===================================================================
+// Implementation of small neural networks
+// ===================================================================
+
+static inline float eval_net(const float *board, const unsigned int n_inp,
+                             const float *w, const float *biases,
+                             const float *logit_w, const float logit_b,
+                             const unsigned int n_dense) {
+  float dense1[FIVE_DENSE_NUM]; // FIVE_DENSE_NUM == SIX_DENSE_NUM
+  float z = logit_b;
+
+  for (unsigned int d1 = 0; d1 < n_dense; d1++) {
+    float h = biases[d1];
+    const float *wd = w + d1 * n_inp;
+    for (unsigned int fl = 0; fl < n_inp; fl++)
+      h += board[fl] * wd[fl];
+    RELU(h);
+    dense1[d1] = h;
+  }
+  for (unsigned int d1 = 0; d1 < n_dense; d1++)
+    z += dense1[d1] * logit_w[d1];
+
+  return fast_tanh(z);
+}
+
+// ===================================================================
+// Interfaces to the two built-in evaluators
+// ===================================================================
+
+static inline float five_nn_fn(tak_state_p state) {
+  static float cur_board[FIVE_INP_NUM] NN_ALIGNED;
+
+  nn_prepare_input(cur_board, state);
+  return eval_net(cur_board, FIVE_INP_NUM, &five_dense_weights[0][0],
+                  five_dense_biases, five_logit_weights, five_logit_bias,
+                  FIVE_DENSE_NUM);
+}
+
+static inline float six_nn_fn(tak_state_p state) {
+  static float cur_board[SIX_INP_NUM] NN_ALIGNED;
 
-typedef float (*nn_function)(tak_state_p);
+  nn_prepare_input(cur_board, state);
+  return eval_net(cur_board, SIX_INP_NUM, &six_dense_weights[0][0],
+                  six_dense_biases, six_logit_weights, six_logit_bias,
+                  SIX_DENSE_NUM);
+}
 
-float nn1986_evaluate_black_win(tak_state_p state);
-float nn2690_evaluate_black_win(tak_state_p state);
+#endif
-- 
cgit v1.3.1