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

1 files changed, 0 insertions, 124 deletions

diff --git a/include/nn.c b/include/nn.c
deleted file mode 100644
index a837a59..0000000
--- a/include/nn.c
+++ /dev/null
@@ -1,124 +0,0 @@
-/*
-  This file is part of ct.
-
-  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
-  the Free Software Foundation, either version 3 of the License, or
-  (at your option) any later version.
-
-  This program is distributed in the hope that it will be useful, but
-  WITHOUT ANY WARRANTY; without even the implied warranty of
-  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
-  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/>.
-*/
-
-#include "nn.h"
-#include "tak.h"
-#include "weights_5.h"
-#include "weights_6.h"
-
-// ===================================================================
-// Helpers
-// ===================================================================
-
-#define RELU(x) ((x) = ((x) < 0) ? 0 : (x))
-
-static inline void prepare_input(float *cur_board, tak_state_p stat);
-
-// ===================================================================
-// Implementation of small neural networks
-// ===================================================================
-
-float nn1986_evaluate_black_win(tak_state_p state) {
-  static float cur_board[FIVE_INP_NUM];
-  static float dense1[FIVE_DENSE_NUM];
-  static float output[2];
-
-  prepare_input(cur_board, state);
-  /* ------------------ *
-   *  First dense layer *
-   * ------------------ */
-  for (unsigned int d1 = 0; d1 < FIVE_DENSE_NUM; d1++) {
-    dense1[d1] = five_dense_biases[d1];
-    for (unsigned int fl = 0; fl < FIVE_INP_NUM; fl++) {
-      dense1[d1] += cur_board[fl] * five_dense_weights[d1][fl];
-    }
-    RELU(dense1[d1]);
-  }
-  /* ------------- *
-   *  Output layer *
-   * ------------- */
-  for (uint8_t k = 0; k < 2; k++) {
-    output[k] = five_output_bias[k];
-    for (unsigned int d2 = 0; d2 < FIVE_DENSE_NUM; d2++) {
-      output[k] += dense1[d2] * five_output_weights[k][d2];
-    }
-    RELU(output[k]);
-  }
-  const float norm = output[0] + output[1];
-  return (2 * output[0] / norm) - 1;
-}
-
-float nn2690_evaluate_black_win(tak_state_p state) {
-  static float cur_board[SIX_INP_NUM];
-  static float dense1[SIX_DENSE_NUM];
-  static float output[2];
-
-  prepare_input(cur_board, state);
-  /* ------------------ *
-   *  First dense layer *
-   * ------------------ */
-  for (unsigned int d1 = 0; d1 < SIX_DENSE_NUM; d1++) {
-    dense1[d1] = six_dense_biases[d1];
-    for (unsigned int fl = 0; fl < SIX_INP_NUM; fl++) {
-      dense1[d1] += cur_board[fl] * six_dense_weights[d1][fl];
-    }
-    RELU(dense1[d1]);
-  }
-  /* ------------- *
-   *  Output layer *
-   * ------------- */
-  for (uint8_t k = 0; k < 2; k++) {
-    output[k] = six_output_bias[k];
-    for (unsigned int d2 = 0; d2 < SIX_DENSE_NUM; d2++) {
-      output[k] += dense1[d2] * six_output_weights[k][d2];
-    }
-    RELU(output[k]);
-  }
-  const float norm = output[0] + output[1];
-  return (2 * output[0] / norm) - 1;
-}
-
-// ===================================================================
-// Helper implementation
-// ===================================================================
-
-static inline void prepare_input(float *cur_board, tak_state_p state) {
-  /* --------------- *
-   *  Populate input *
-   * --------------- */
-  for (unsigned int y = 0; y < state->board_size; y++) {
-    for (unsigned int x = 0; x < state->board_size; x++) {
-      const unsigned int loc = x + y * state->board_size;
-      const unsigned int count = COUNT_AT(state, loc);
-      float lookup = 0;
-      if (count > 0) {
-        if (STONE_AT(state, loc) == STONE_STANDING) {
-          lookup = (state->colours[loc] & 1) ? +0.25 : -0.25;
-        } else if (STONE_AT(state, loc) == STONE_CAPSTONE) {
-          lookup = (state->colours[loc] & 1) ? +1.00 : -1.00;
-        } else {
-          lookup = (state->colours[loc] & 1) ? +0.50 : -0.50;
-        }
-      }
-      cur_board[3 + loc] = lookup;
-    }
-  }
-  // Add input of flat counts and ply parity
-  cur_board[0] = (state->ply & 1) ? 1 : -1;
-  cur_board[1] = (float)(state->white_count & 127) / 21.0;
-  cur_board[2] = (float)(state->black_count & 127) / 21.0;
-}