diff options
Diffstat (limited to 'include/minimax_cnn1986.c')
| -rw-r--r-- | include/minimax_cnn1986.c | 362 |
1 files changed, 362 insertions, 0 deletions
diff --git a/include/minimax_cnn1986.c b/include/minimax_cnn1986.c new file mode 100644 index 0000000..ce950ba --- /dev/null +++ b/include/minimax_cnn1986.c @@ -0,0 +1,362 @@ +#include "minimax_cnn1986.h" + +// =================================================================== +// Globals +// =================================================================== + +const float infty = 3.0; +char ct1986_ptn[9]; +void (*ct1986_display_progress)(const uint8_t); + +// =================================================================== +// Implementation of a small convolutional neural network +// =================================================================== + +static float flattened[CONV_NUM+2]; +static float dense1[DENSE1_NUM]; +static float dense2[DENSE2_NUM]; + +#ifndef DETERMINISTIC +union u_f { + uint32_t u; + float f; +}; + +static uint32_t state = 1; +static union u_f fudge; + +#define DOXORSHIFT { \ + state ^= state << 13; \ + state ^= state >> 17; \ + state ^= state << 5; \ + fudge.u = 0x3f800000 | state >> 10; \ + fudge.f = (fudge.f - 1.5) * 0.01; \ +} +#endif + +#define RELU(x) ((x) = ((x)<0)?0:(x)) + +float +ct1986_evaluate_black_win(void) { + /* ------------------ * + * Convolution layer * + * ------------------ */ + // for each kernel + for (uint8_t kern = 0; kern < KERN_NUM; kern++) { + // the stride is 1, march across the board + for (uint8_t bx = 0; bx < KERN_OSIZE; bx++) { + for (uint8_t by = 0; by < KERN_OSIZE; by++) { + flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)] = + conv2d_biases[kern]; + // Compute the convolution for this position + for (uint8_t ky = 0; ky < KERN_SIZE; ky++) { + for (uint8_t kx = 0; kx < KERN_SIZE; kx++) { + for (uint8_t c = 0; c < KERN_CHAN; c++) { + // Where we are on the board + const uint8_t loc = kx+bx+(ky+by)*5; + // Look up what's on the board at this location, and + // multiply it. For c=0 we have to do some extra work + float lookup = 0; + if (COUNT_AT(loc)>c) { + if (c==0) { + if (STONE_AT(loc) == STONE_STANDING) { + lookup = (colours[loc] & 1) ? +0.25 : -0.25; + } else if (STONE_AT(loc) == STONE_CAPSTONE) { + lookup = (colours[loc] & 1) ? +1.00 : -1.00; + } else { + lookup = (colours[loc] & 1) ? +0.50 : -0.50; + } + } else { + lookup = (colours[loc] & (1<<c)) ? +0.50 : -0.50; + } + } + flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)] + += lookup*conv2d_weights[kern][ky][kx][c]; + } + } + } + RELU(flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)]); + } + } + } + // Add input of flat counts + flattened[CONV_NUM] = (float)(white_count & 127)/21.0; + flattened[CONV_NUM+1] = (float)(black_count & 127)/21.0; + /* ------------------ * + * First dense layer * + * ------------------ */ + for (uint8_t d1 = 0; d1 < DENSE1_NUM; d1++) { + dense1[d1] = dense1_biases[d1]; + for (uint8_t fl = 0; fl < CONV_NUM+2; fl++) { + dense1[d1] += flattened[fl]*dense1_weights[d1][fl]; + } + RELU(dense1[d1]); + } + /* ------------------- * + * Second dense layer * + * ------------------- */ + for (uint8_t d2 = 0; d2 < DENSE2_NUM; d2++) { + dense2[d2] = dense2_biases[d2]; + for (uint8_t d1 = 0; d1 < DENSE1_NUM; d1++) { + dense2[d2] += dense1[d1]*dense2_weights[d2][d1]; + } + RELU(dense2[d2]); + } + /* ------------- * + * Output layer * + * ------------- */ + float output = output_bias; + for (uint8_t d2 = 0; d2 < DENSE2_NUM; d2++) { + output += dense2[d2]*output_weights[d2]; + } + // Truncated Pade approximant of logistic function + output = (12.0+output+50.0*output/(output*output+10.0))/24.0; +#ifndef DETERMINISTIC + DOXORSHIFT; + output += fudge.f; +#endif + if (output > 1.0) return 1.0; + else if (output < 0.0) return 0.0; + + return output; +} + +// =================================================================== +// α-β minimax using the above evaluator +// =================================================================== + +static void +previous_ply(void) { + if (ply>0) ply--; + if (ply == 1) { + current_colour = C_WHITE; + } else { + if (current_colour == C_BLACK) current_colour = C_WHITE; + else current_colour = C_BLACK; + } +} + +static float val; +static enum WIN_TYPE w; + +#define WIN_EVALUATE_OR_RECURSE(store) { \ + w = 0xFF; \ + if (ply >= 2*5 - 2) w = check_win(); \ + if (w < 0xFF) { \ + /* Somebody won, assign weights accordingly */ \ + if (min == 0) { \ + if (w == WIN_ROAD_BLACK || w == WIN_FLAT_BLACK) \ + val = infty; \ + else val = -infty; \ + } else { \ + if (w == WIN_ROAD_WHITE || w == WIN_FLAT_WHITE) \ + val = -infty; \ + else val = infty; \ + } \ + } else if (cur_depth == max_depth) { \ + /* We're at the bottom, evaluate */ \ + val = ct1986_evaluate_black_win(); \ + if ((ply & 1) == 0) val = val - 1.0; \ + } else { \ + /* We're not at the bottom, recurse first */ \ + next_ply(); \ + val = ct1986_minimax(cur_depth + 1, max_depth, 1-min, alpha, beta); \ + previous_ply(); \ + } \ + /* Update the optimal value */ \ + if (((min > 0) && (val <= optimal)) \ + || ((min == 0) && (val >= optimal))) { \ + optimal = val; \ + if (cur_depth == 0) (store); \ + } \ + /* Update alpha and beta */ \ + if (min) { \ + if (optimal < beta) beta = optimal; \ + } else { \ + if (optimal > alpha) alpha = optimal; \ + } \ +} + +float +ct1986_minimax(const uint8_t cur_depth, const uint8_t max_depth, + const uint8_t min, float alpha, float beta) { + + enum E_RESULT r; + const uint8_t black = (ply & 1), + material = (black) ? black_count : white_count, + flat = material & 127, + cap = (ply > 2 && (material & 128)), + standing = (ply > 2 && (material & 127)); + + // 1.0 is a `certain' black win, -1.0 is a `certain' white win. + float optimal = (min) ? infty : -infty; + + // Step across the board + for (uint8_t row = 0; row < 5; row++) { + for (uint8_t col = 0; col < 5; col++) { + // Try all valid actions for this square. Is it empty? + const uint8_t loc = THE_COORDS(col, row); + const uint8_t count = (COUNT_AT(loc) > 5) ? 5 : COUNT_AT(loc); + // Only try moves after CPS + if (count && ((colours[loc] & 1) == current_colour) && ply>2) { + // There are stones, can we move them in a given direction? + // I'm not a huge fan of looping through enums, but it's + // better than manually unrolling this. Sufficiently smart + // compilers? + + uint16_t colours_backup[5]; + uint8_t celldat_backup[5], drops[5]; // we only use 4, the + // fifth is to skip a + // bounds check at (*) + // Back up the row of the board + for (uint8_t y = 0; y < 5; y++) { + colours_backup[y] = colours[THE_COORDS(col, y)]; + celldat_backup[y] = celldat[THE_COORDS(col, y)]; + } + for (enum MOVE_DIRECTION dir = M_UP; dir <= M_RIGHT; dir++) { + // Back-up the column once we start looking horizontally + if (dir == M_LEFT) { + for (uint8_t x = 0; x < 5; x++) { + colours_backup[x] = colours[THE_COORDS(x, row)]; + celldat_backup[x] = celldat[THE_COORDS(x, row)]; + } + } + /* + * We don't do anything terribly efficient or smart here, + * just try all the ordered partitions of num ∈ {1,…,count} + * that will fit on the board in the current direction. + * Recall that count = max(stack height, carry limit). + */ + uint8_t upper; + switch (dir) { + case M_UP: { + upper = (4-row > count) ? count : 4-row; + break; + } + case M_DOWN: { + upper = (row > count) ? count : row; + break; + } + case M_LEFT: { + upper = (col > count) ? count : col; + break; + } + case M_RIGHT: { + upper = (4-col > count) ? count : 4-col; + break; + } + } + uint8_t gaps, t, idx, mask; + for (uint8_t num = 1; num <= count; num++) { + for (uint8_t steps = 1; steps <= upper && steps <= num; steps++) { + gaps = 0b00000111 >> (4-steps); + do { + // Translate to a drop sequence + drops[0] = 1; mask = 1; idx = 0; + for (uint8_t d = 0; d + 1 < num; d++) { + if (gaps & mask) { + idx++; + drops[idx] = 1; // (*) we don't need to bounds check + } else { + drops[idx] += 1; + } + mask <<= 1; + } + // Try it, and manually check for win if it's valid + r = try_move(loc, dir, steps, drops); + if (r == ACT_OK) { + // First check for wins, if we're at the bottom + // evaluate, otherwise recurse + WIN_EVALUATE_OR_RECURSE({ + // If we did update the optimal value, store + // this move + generate_move(loc, dir, steps, drops, ct1986_ptn); + }); + // Reset the board data + if (dir <= M_DOWN) { + for (uint8_t y = 0; y < 5; y++) { + colours[THE_COORDS(col, y)] = colours_backup[y]; + celldat[THE_COORDS(col, y)] = celldat_backup[y]; + } + } else { + for (uint8_t x = 0; x < 5; x++) { + colours[THE_COORDS(x, row)] = colours_backup[x]; + celldat[THE_COORDS(x, row)] = celldat_backup[x]; + } + } + } + // Prune + if (alpha >= beta) return optimal; + /* + * With thanks to + * https://graphics.stanford.edu/~seander/bithacks.html#NextBitPermutation + * we have the following magic to generate the next + * permutation of steps-many set bits + */ + t = (gaps | (gaps - 1)); + gaps = (t + 1) | (((~t & -~t) - 1) >> (__builtin_ctz(gaps) + 1)); + } while (gaps && (gaps + 1 <= (1<<(num-1)))); + } + } + } + } else if (material && count == 0) { + // Empty square, try placements + + if (flat) { + // Generate the placement + if (black) black_count--; + else white_count--; + colours[loc] = current_colour; + celldat[loc] = NUM_INC | STONE_FLAT; + WIN_EVALUATE_OR_RECURSE({ + // If we did update the optimal value, store + generate_place(loc, STONE_FLAT, ct1986_ptn); + }); + // Reset the state + celldat[loc] = 0; + if (black) black_count++; + else white_count++; + // Prune + if (alpha >= beta) return optimal; + + // Do the same for walls, can't happen without flats + if (standing) { + if (black) black_count--; + else white_count--; + colours[loc] = current_colour; + celldat[loc] = NUM_INC | STONE_STANDING; + WIN_EVALUATE_OR_RECURSE({ + generate_place(loc, STONE_STANDING, ct1986_ptn); + }); + celldat[loc] = 0; + if (black) black_count++; + else white_count++; + if (alpha >= beta) return optimal; + } + } + + // and for caps + if (cap) { + if (black) black_count &= 127; + else white_count &= 127; + colours[loc] = current_colour; + celldat[loc] = NUM_INC | STONE_CAPSTONE; + WIN_EVALUATE_OR_RECURSE({ + generate_place(loc, STONE_CAPSTONE, ct1986_ptn); + }); + celldat[loc] = 0; + if (black) black_count |= 128; + else white_count |= 128; + if (alpha >= beta) return optimal; + } + } + ct1986_display_progress(cur_depth); + } + } + return optimal; +} + +inline float +ct1986_generate(const uint8_t max_depth) { + return ct1986_minimax(0, max_depth, (ply & 1) ? 0 : 1, -infty, infty); +} |