Files
cc65-Chess/tests/genbook.c
T
Stefan WesselsandClaude Opus 5 d9d49f7f8e Replace the engine: 0x88 board, make/unmake, alpha-beta search
The old engine was weak for one structural reason. cpu_ScorePieceSubTree
looked like minimax but had a branching factor of one: at each level it
followed only the single reply a static heuristic ranked first, so it
evaluated one predicted line per root move and never saw a refutation the
heuristic missed. Two things compounded it - the root considered only one
move per piece, and the scoring functions returned move desirability rather
than position value, which is not a quantity minimax can validly accumulate
across plies.

What replaces it:

  engine.c  0x88 board, make/unmake against a per-ply state stack, legal
            move generation, and attack queries by ray-cast.  The stored
            attack database is gone - eng_IsAttacked walks outwards from a
            square instead of rebuilding 64 squares' worth of state after
            every trial move.
  eval.c    material and piece-square tables as a pure position value,
            kept as a running total updated incrementally by make/unmake.
  search.c  iterative-deepening negamax with alpha-beta, MVV-LVA and killer
            ordering, and a quiescence search over captures.

board.c becomes the adapter between engine and display; plat.h and the 0-63
tile numbering are untouched, so no platform file needed editing - including
the three that cannot be built here.

Verified: perft exact on all five reference positions to depth 5, and to
depth 6 on the two that publish one. Against the engine it replaces, at that
engine's strongest setting: 6 wins, 0 losses.

Smaller than what it replaces despite doing far more - 31129 bytes total on
the c64 at optspeed against 31543 for the two engines' worth of machinery it
supersedes, and 514 bytes less RAM. All five buildable targets fit.

tests/ carries the harness this was developed against: perft, a game fuzzer,
castling scenarios, an internal match runner, on-target C64 benchmarks, and
a UCI adapter for playing external engines. It stops being gitignored here -
the exclusion was scaffolding added during the rework to keep git status
quiet, and the harness is the evidence for everything above.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-07 12:36:05 -07:00

69 lines
1.7 KiB
C

/*
* genbook.c
* cc65 Chess - test support
*
* Writes an EPD opening book on stdout, for a match runner to start games
* from. The engine is deterministic, so without a book a match is one game
* repeated N times - see the note at the top of tests/uci.c.
*
* The positions are generated the same way tests/match.c generates its own,
* on purpose: a few plies of random legal non-capturing moves, one fixed seed
* per opening. That keeps an external match commensurable with the internal
* one, and it is reproducible without a page of hand-typed FENs to get wrong.
*
* ./genbook [count] [plies] > book.epd
*/
#include <stdio.h>
#include <stdlib.h>
#include "types.h"
#include "globals.h"
#include "engine.h"
#include "testutil.h"
/*-----------------------------------------------------------------------*/
int main(int argc, char **argv)
{
int count = (argc > 1) ? atoi(argv[1]) : 256;
int plies = (argc > 2) ? atoi(argv[2]) : 4;
int i;
for(i = 0; i < count; ++i)
{
char side = SIDE_WHITE, fen[90];
int ply;
srand(1000 + i);
eng_SetStartPosition();
for(ply = 0; ply < plies; ++ply)
{
t_engMove moves[ENG_MAX_MOVES];
t_engUndo undo;
char n = eng_GenMoves(side, moves, ENG_MAX_MOVES);
char tries, done = 0;
// no captures: a book position where one side is already a piece
// up measures the opening set rather than the engines
for(tries = 0; tries < n * 2 && !done; ++tries)
{
char pick = rand() % n;
eng_Make(&moves[pick], &undo);
if(eng_IsAttacked(geKing[side], 1 - side) ||
NONE != (undo.m_captured & PIECE_DATA))
eng_Unmake(&moves[pick], &undo);
else
done = 1;
}
if(!done)
break;
side = 1 - side;
}
test_EngineGetFEN(side, fen);
printf("%s\n", fen);
}
return 0;
}