RGBCube/serenity
1
Fork 0
mirror of https://github.com/RGBCube/serenity synced 2025-07-27 05:17:35 +00:00
Code Issues Activity Actions

ChessEngine: don't store board in non-leaf nodes in MCTS

Browse source 
Also make parameters static so they aren't in every node of the tree
this saves a substantial amount of memory.
This commit is contained in:
Peter Elliott 2021-06-19 17:27:59 -06:00 • committed by Andreas Kling
parent 34433f5dc4
commit 49539abee0
3 changed files with 30 additions and 30 deletions
Download patch file Download diff file Expand all files Collapse all files

3
Userland/Services/ChessEngine/ChessEngine.cpp
View file

@ -42,9 +42,6 @@ void ChessEngine::handle_go(const GoCommand& command)
MCTSTree mcts(m_board);
// FIXME: optimize simulations enough for use.
mcts.set_eval_method(MCTSTree::EvalMethod::Heuristic);
int rounds = 0;
while (elapsed_time.elapsed() <= command.movetime.value()) {
mcts.do_round();

37
Userland/Services/ChessEngine/MCTSTree.cpp
View file

@ -8,13 +8,12 @@
#include <AK/String.h>
#include <stdlib.h>
MCTSTree::MCTSTree(const Chess::Board& board, double exploration_parameter, MCTSTree* parent)
MCTSTree::MCTSTree(const Chess::Board& board, MCTSTree* parent)
: m_parent(parent)
, m_exploration_parameter(exploration_parameter)
, m_board(board)
, m_board(make<Chess::Board>(board))
, m_last_move(board.last_move())
, m_turn(board.turn())
{
if (m_parent)
m_eval_method = m_parent->eval_method();
}
MCTSTree& MCTSTree::select_leaf()
@ -25,7 +24,7 @@ MCTSTree& MCTSTree::select_leaf()
MCTSTree* node = nullptr;
double max_uct = -double(INFINITY);
for (auto& child : m_children) {
double uct = child.uct(m_board.turn());
double uct = child.uct(m_turn);
if (uct >= max_uct) {
max_uct = uct;
node = &child;
@ -40,13 +39,15 @@ MCTSTree& MCTSTree::expand()
VERIFY(!expanded() || m_children.size() == 0);
if (!m_moves_generated) {
m_board.generate_moves([&](Chess::Move move) {
Chess::Board clone = m_board;
m_board->generate_moves([&](Chess::Move move) {
Chess::Board clone = *m_board;
clone.apply_move(move);
m_children.append(make<MCTSTree>(clone, m_exploration_parameter, this));
m_children.append(make<MCTSTree>(clone, this));
return IterationDecision::Continue;
});
m_moves_generated = true;
if (m_children.size() != 0)
m_board = nullptr; // Release the board to save memory.
}
if (m_children.size() == 0) {
@ -63,8 +64,7 @@ MCTSTree& MCTSTree::expand()
int MCTSTree::simulate_game() const
{
VERIFY_NOT_REACHED();
Chess::Board clone = m_board;
Chess::Board clone = *m_board;
while (!clone.game_finished()) {
clone.apply_move(clone.random_move());
}
@ -73,10 +73,10 @@ int MCTSTree::simulate_game() const
int MCTSTree::heuristic() const
{
if (m_board.game_finished())
return m_board.game_score();
if (m_board->game_finished())
return m_board->game_score();
double winchance = max(min(double(m_board.material_imbalance()) / 6, 1.0), -1.0);
double winchance = max(min(double(m_board->material_imbalance()) / 6, 1.0), -1.0);
double random = double(rand()) / RAND_MAX;
if (winchance >= random)
@ -101,7 +101,7 @@ void MCTSTree::do_round()
auto& node = select_leaf().expand();
int result;
if (m_eval_method == EvalMethod::Simulation) {
if constexpr (s_eval_method == EvalMethod::Simulation) {
result = node.simulate_game();
} else {
result = node.heuristic();
@ -111,7 +111,7 @@ void MCTSTree::do_round()
Chess::Move MCTSTree::best_move() const
{
int score_multiplier = (m_board.turn() == Chess::Color::White) ? 1 : -1;
int score_multiplier = (m_turn == Chess::Color::White) ? 1 : -1;
Chess::Move best_move = { { 0, 0 }, { 0, 0 } };
double best_score = -double(INFINITY);
@ -119,8 +119,7 @@ Chess::Move MCTSTree::best_move() const
for (auto& node : m_children) {
double node_score = node.expected_value() * score_multiplier;
if (node_score >= best_score) {
// The best move is the last move made in the child.
best_move = node.m_board.moves()[node.m_board.moves().size() - 1];
best_move = node.m_last_move.value();
best_score = node_score;
}
}
@ -143,7 +142,7 @@ double MCTSTree::uct(Chess::Color color) const
// Fun fact: Szepesvári was my data structures professor.
double expected = expected_value() * ((color == Chess::Color::White) ? 1 : -1);
return expected + m_exploration_parameter * sqrt(log(m_parent->m_simulations) / m_simulations);
return expected + s_exploration_parameter * sqrt(log(m_parent->m_simulations) / m_simulations);
}
bool MCTSTree::expanded() const

20
Userland/Services/ChessEngine/MCTSTree.h
View file

@ -8,6 +8,7 @@
#include <AK/Function.h>
#include <AK/NonnullOwnPtrVector.h>
#include <AK/OwnPtr.h>
#include <LibChess/Chess.h>
#include <math.h>
@ -18,7 +19,7 @@ public:
Heuristic,
};
MCTSTree(const Chess::Board& board, double exploration_parameter = sqrt(2), MCTSTree* parent = nullptr);
MCTSTree(const Chess::Board& board, MCTSTree* parent = nullptr);
MCTSTree& select_leaf();
MCTSTree& expand();
@ -32,16 +33,19 @@ public:
double uct(Chess::Color color) const;
bool expanded() const;
EvalMethod eval_method() const { return m_eval_method; }
void set_eval_method(EvalMethod method) { m_eval_method = method; }
private:
// While static parameters are less configurable, they don't take up any
// memory in the tree, which I believe to be a worthy tradeoff.
static constexpr double s_exploration_parameter { sqrt(2) };
// FIXME: Optimize simulations enough for use.
static constexpr EvalMethod s_eval_method { EvalMethod::Heuristic };
NonnullOwnPtrVector<MCTSTree> m_children;
MCTSTree* m_parent { nullptr };
int m_white_points { 0 };
int m_simulations { 0 };
bool m_moves_generated { false };
double m_exploration_parameter;
EvalMethod m_eval_method { EvalMethod::Simulation };
Chess::Board m_board;
OwnPtr<Chess::Board> m_board;
Optional<Chess::Move> m_last_move;
Chess::Color m_turn : 2;
bool m_moves_generated : 1 { false };
};