def _traverse_from(self, grid: Grid, current_color: Color) -> bool:
self._nodes_for_backprop.append((grid.state, current_color))
node_info: NodeInfo = self._tree[(grid.state, current_color)]
if not node_info.is_leaf:
move = self._select_best_child_ucb(grid, current_color)
return self._traverse_from(
grid.grid_after_move(current_color, move),
Color(1 - current_color))
elif len(grid.available_moves) == 0:
return False
elif node_info.visits == 0:
return self._rollout_from(grid, current_color)
else:
self._tree[(grid.state,
current_color)] = NodeInfo(wins=node_info.wins,
visits=node_info.visits,
is_leaf=False)
for move in grid.available_moves:
new_grid = grid.grid_after_move(current_color, move)
if (new_grid.state,
Color(1 - current_color)) not in self._tree.keys():
self._tree[(new_grid.state,
1 - current_color)] = NodeInfo(wins=0,
visits=0,
is_leaf=True)
move = random.choice(grid.available_moves)
return self._traverse_from(
grid.grid_after_move(current_color, move),
Color(1 - current_color))
def test_move(self):
g = Grid()
g.move(Color.RED, 0)
state = g.state.cols_first
self.assertEqual(state[0][0], Color.RED)
self.assertIsNone(state[0][1])
def _minmax(self, grid: Grid, depth: int, alpha: int, beta: int,
color: Color, last_move: Union[int, None]) -> Tuple[int, int]:
"""Returns (best available value, move towards best value)"""
computed = self._transposition_table.get(grid.state, None)
if computed is not None and computed[0] >= depth: return computed[1:]
if self._deadline.is_set() or depth == 0 or \
(last_move is not None and self._judge.is_over_after_move_in_col(grid.state, last_move)):
return self._evaluate(grid.state, self._judge), 0
value = -INF if color == MAX_COLOR else INF
best_move = None
for move in grid.available_moves:
child_value, _ = self._minmax(grid.grid_after_move(color, move),
depth - 1, alpha, beta,
Color(1 - color), move)
if color == MAX_COLOR and child_value > value:
best_move = move
value = child_value
alpha = max(alpha, value)
elif color == MIN_COLOR and child_value < value:
best_move = move
value = child_value
beta = min(beta, value)
if alpha >= beta: break
self._transposition_table[grid.state] = (depth, value, best_move)
return value, best_move
def create_game(first_player: Player,
second_player: Player,
first_color: Color = Color.RED) -> Game:
return Game(grid=Grid(),
judge=judge,
first_player=first_player,
second_player=second_player,
first_color=first_color)
def _rollout_from(self,
grid: Grid,
color: Color,
last_col: Union[int, None] = None) -> bool:
if len(grid.available_moves) == 0 or (last_col is None and self._judge.is_over(grid.state)) or \
(last_col is not None and self._judge.is_over_after_move_in_col(grid.state, last_col)):
return color != self._color
move = random.choice(grid.available_moves)
has_won = self._rollout_from(grid.grid_after_move(color, move),
Color(1 - color), move)
return has_won
def make_move_in_state(self, state: State) -> int:
grid = Grid.from_state(state)
assert len(grid.available_moves) > 0, 'No move available'
finishing_move = self._finishing_move_in(grid)
if finishing_move is not None: return finishing_move
if (state, self._color) not in self._tree.keys():
self._tree[(state, self._color)] = NodeInfo(visits=0,
is_leaf=True,
wins=0)
self._compute(grid)
return self._pick_most_visited_child_of(grid)
def test_middle_finishing(self):
judge = Judge()
g = Grid(ncols=4, nrows=2)
g.move(Color.RED, 0)
g.move(Color.RED, 1)
g.move(Color.RED, 3)
g.move(Color.RED, 2)
self.assertTrue(judge.is_over(g.state))
self.assertTrue(judge.is_over_after_move_in_col(g.state, 2))
def _select_best_child(self, parent: Grid, map_node_info: callable,
current_color: Color) -> int:
"""map_node_info should take parent's node_info and child's and return a positive float"""
assert len(parent.available_moves) > 0, 'A leaf actually'
parent_info = self._tree[(parent.state, current_color)]
best_move, best_result = None, -1
for move in parent.available_moves:
child_info = self._tree[(parent.grid_after_move(
current_color, move).state, Color(1 - current_color))]
temp_result = map_node_info(parent_info, child_info)
if temp_result > best_result:
best_result, best_move = temp_result, move
return best_move
def test_initial_state_rows_first(self):
state = Grid(ncols=7, nrows=6).state.rows_first
self.assertEqual(len(state), 6)
self.assertEqual(len(state[0]), 7)
def test_size(self):
g = Grid(ncols=12, nrows=2)
self.assertEqual(g.ncols, 12)
self.assertEqual(g.nrows, 2)
def test_vertical_win(self):
judge = Judge()
g = Grid(ncols=2, nrows=4)
g.move(Color.RED, 0)
g.move(Color.BLACK, 1)
g.move(Color.RED, 0)
g.move(Color.BLACK, 1)
g.move(Color.RED, 0)
g.move(Color.BLACK, 1)
self.assertFalse(judge.is_over(g.state))
g.move(Color.RED, 0)
self.assertTrue(judge.is_over(g.state))
self.assertTrue(judge.is_over_vertical(g.state))
self.assertFalse(judge.is_over_horizontal(g.state))
def _finishing_move_in(self, grid: Grid) -> Union[int, None]:
for move in grid.available_moves:
after_move = grid.grid_after_move(self._color, move)
if self._judge.is_over_after_move_in_col(after_move.state, move):
return move
def test_state(self):
g = Grid(ncols=5, nrows=4)
g.move(Color.RED, 0)
g.move(Color.BLACK, 1)
g.move(Color.RED, 1)
g.move(Color.BLACK, 2)
g.move(Color.RED, 4)
g.move(Color.BLACK, 1)
state = g.state.rows_first
# 3 - - - - -
# 2 - B - - -
# 1 - R - - -
# 0 R B B - R
#
# 0 1 2 3 4
self.assertSequenceEqual([None, None, None, None, None], state[3])
self.assertSequenceEqual([None, Color.BLACK, None, None, None],
state[2])
self.assertSequenceEqual([None, Color.RED, None, None, None], state[1])
self.assertSequenceEqual(
[Color.RED, Color.BLACK, Color.BLACK, None, Color.RED], state[0])
def test_rightup_win(self):
judge = Judge()
g = Grid(ncols=5, nrows=4)
g.move(Color.RED, 1)
g.move(Color.BLACK, 0)
g.move(Color.RED, 2)
g.move(Color.BLACK, 1)
g.move(Color.RED, 3)
g.move(Color.BLACK, 2)
g.move(Color.RED, 3)
g.move(Color.BLACK, 2)
g.move(Color.RED, 3)
# 3 - - - B -
# 2 - - B R -
# 1 - B B R -
# 0 B R R R -
# 0 1 2 3 4
self.assertFalse(judge.is_over(g.state))
g.move(Color.BLACK, 3)
self.assertTrue(judge.is_over(g.state))
self.assertTrue(judge.is_over_rightup(g.state))
self.assertFalse(judge.is_over_rightdown(g.state))
def test_move_outside(self):
g = Grid()
with self.assertRaises(AssertionError):
g.move(Color.RED, 7)
def make_move_in_state(self, state: State) -> int:
self._deadline.clear()
Timer(self._timeout, lambda: self._deadline.set()).start()
return self._iterative_deepening(Grid.from_state(state))
(last_move is not None and self._judge.is_over_after_move_in_col(grid.state, last_move)):
return self._evaluate(grid.state, self._judge), 0
value = -INF if color == MAX_COLOR else INF
best_move = None
for move in grid.available_moves:
child_value, _ = self._minmax(grid.grid_after_move(color, move),
depth - 1, alpha, beta,
Color(1 - color), move)
if color == MAX_COLOR and child_value > value:
best_move = move
value = child_value
alpha = max(alpha, value)
elif color == MIN_COLOR and child_value < value:
best_move = move
value = child_value
beta = min(beta, value)
if alpha >= beta: break
self._transposition_table[grid.state] = (depth, value, best_move)
return value, best_move
if __name__ == '__main__':
game = Game(Grid(), Judge(),
MinmaxPlayer(Color.RED, Judge(), Evaluator(), 4, 30),
MinmaxPlayer(Color.BLACK, Judge(), Evaluator(), 6, 18))
print(game.play())
def test_from_coords(self):
judge = Judge()
g = Grid(ncols=5, nrows=4)
g.move(Color.RED, 2)
g.move(Color.BLACK, 3)
g.move(Color.RED, 1)
g.move(Color.BLACK, 2)
g.move(Color.RED, 0)
g.move(Color.BLACK, 1)
g.move(Color.RED, 0)
g.move(Color.BLACK, 1)
g.move(Color.RED, 0)
g.move(Color.BLACK, 0)
# 3 B - - - -
# 2 R B - - -
# 1 R B B - -
# 0 R R R B -
# 0 1 2 3 4
self.assertFalse(judge.is_over_from(g.state, 1, 2))
self.assertFalse(judge.is_over_from(g.state, 2, 1))
self.assertTrue(judge.is_over_from(g.state, 0, 3))
self.assertTrue(judge.is_over_from(g.state, 3, 0))
def test_horizontal_win(self):
judge = Judge()
g = Grid(ncols=4, nrows=4)
g.move(Color.RED, 0)
g.move(Color.BLACK, 0)
g.move(Color.RED, 1)
g.move(Color.BLACK, 1)
g.move(Color.RED, 2)
g.move(Color.BLACK, 2)
self.assertFalse(judge.is_over(g.state))
g.move(Color.RED, 3)
self.assertTrue(judge.is_over(g.state))
self.assertTrue(judge.is_over_horizontal(g.state))
self.assertFalse(judge.is_over_rightdown(g.state))
def test_rightdown_win(self):
judge = Judge()
g = Grid(ncols=5, nrows=4)
g.move(Color.RED, 2)
g.move(Color.BLACK, 3)
g.move(Color.RED, 1)
g.move(Color.BLACK, 2)
g.move(Color.RED, 0)
g.move(Color.BLACK, 1)
g.move(Color.RED, 0)
g.move(Color.BLACK, 1)
g.move(Color.RED, 0)
# 3 B - - - -
# 2 R B - - -
# 1 R B B - -
# 0 R R R B -
# 0 1 2 3 4
self.assertFalse(judge.is_over(g.state))
g.move(Color.BLACK, 0)
self.assertTrue(judge.is_over(g.state))
self.assertTrue(judge.is_over_rightdown(g.state))
self.assertFalse(judge.is_over_vertical(g.state))
def test_overflow(self):
g = Grid()
for _ in range(6):
g.move(Color.RED, 2)
with self.assertRaises(AssertionError):
g.move(Color.RED, 2)
move = random.choice(grid.available_moves)
return self._traverse_from(
grid.grid_after_move(current_color, move),
Color(1 - current_color))
def _rollout_from(self,
grid: Grid,
color: Color,
last_col: Union[int, None] = None) -> bool:
if len(grid.available_moves) == 0 or (last_col is None and self._judge.is_over(grid.state)) or \
(last_col is not None and self._judge.is_over_after_move_in_col(grid.state, last_col)):
return color != self._color
move = random.choice(grid.available_moves)
has_won = self._rollout_from(grid.grid_after_move(color, move),
Color(1 - color), move)
return has_won
def _finishing_move_in(self, grid: Grid) -> Union[int, None]:
for move in grid.available_moves:
after_move = grid.grid_after_move(self._color, move)
if self._judge.is_over_after_move_in_col(after_move.state, move):
return move
if __name__ == '__main__':
game = Game(Grid(), Judge(), MCTSPlayer(Color.RED, Judge(), 2, 1000),
MCTSPlayer(Color.BLACK, Judge(), 5, 1000))
print(game.play())
def test_empty_grid(self):
Grid(ncols=7, nrows=6)
def test_from_last_in_empty(self):
judge = Judge()
g = Grid(ncols=2, nrows=2)
with self.assertRaises(AssertionError):
judge.is_over_after_move_in_col(g.state, 1)