hongshaorou/Tianshou
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix the copy bug in check_global_isomorphous; refactor code to eliminate side effect

Browse Source
This commit is contained in:
Dong Yan 2017-12-19 22:57:38 +08:00
parent 83f9e19fa5
commit f8a70183b6
3 changed files with 67 additions and 76 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
AlphaGo/go.py
View File

@ -72,18 +72,14 @@ class Go:
self.game.board[self.game._flatten(vertex)] = utils.EMPTY self.game.board[self.game._flatten(vertex)] = utils.EMPTY
return True return True
def _check_global_isomorphous(self, color, vertex): def _check_global_isomorphous(self, history_boards, current_board, color, vertex):
##backup repeat = False
_board = copy.copy(self.game.board) next_board = copy.copy(current_board)
self.game.board[self.game._flatten(vertex)] = color next_board[self.game._flatten(vertex)] = color
self._process_board(color, vertex) self._process_board(next_board, color, vertex)
if self.game.board in self.game.history: if next_board in history_boards:
res = True repeat = True
else: return repeat
res = False
self.game.board = _board
return res
def _in_board(self, vertex): def _in_board(self, vertex):
x, y = vertex x, y = vertex
@ -101,38 +97,38 @@ class Go:
nei.append((_x, _y)) nei.append((_x, _y))
return nei return nei
def _process_board(self, color, vertex): def _process_board(self, current_board, color, vertex):
nei = self._neighbor(vertex) nei = self._neighbor(vertex)
for n in nei: for n in nei:
if self.game.board[self.game._flatten(n)] == utils.another_color(color): if current_board[self.game._flatten(n)] == utils.another_color(color):
can_kill, block = self._find_block(n) can_kill, block = self._find_block(n)
if can_kill: if can_kill:
for b in block: for b in block:
self.game.board[self.game._flatten(b)] = utils.EMPTY current_board[self.game._flatten(b)] = utils.EMPTY
def is_valid(self, color, vertex): def is_valid(self, history_boards, current_board, color, vertex):
### in board ### in board
if not self._in_board(vertex): if not self._in_board(vertex):
return False return False
### already have stone ### already have stone
if not self.game.board[self.game._flatten(vertex)] == utils.EMPTY: if not current_board[self.game._flatten(vertex)] == utils.EMPTY:
return False return False
### check if it is qi ### check if it is qi
if not self._is_qi(color, vertex): if not self._is_qi(color, vertex):
return False return False
if self._check_global_isomorphous(color, vertex): if self._check_global_isomorphous(history_boards, current_board, color, vertex):
return False return False
return True return True
def do_move(self, color, vertex): def do_move(self, color, vertex):
if not self.is_valid(color, vertex): if not self.is_valid(self.game.history, self.game.board, color, vertex):
return False return False
self.game.board[self.game._flatten(vertex)] = color self.game.board[self.game._flatten(vertex)] = color
self._process_board(color, vertex) self._process_board(self.game.board, color, vertex)
self.game.history.append(copy.copy(self.game.board)) self.game.history.append(copy.copy(self.game.board))
self.game.latest_boards.append(copy.copy(self.game.board)) self.game.latest_boards.append(copy.copy(self.game.board))
return True return True

104
AlphaGo/strategy.py
View File

@ -19,52 +19,47 @@ class GoEnv:
self.simulate_board = [utils.EMPTY] * (self.game.size ** 2) self.simulate_board = [utils.EMPTY] * (self.game.size ** 2)
self.simulate_latest_boards = deque(maxlen=8) self.simulate_latest_boards = deque(maxlen=8)
def _find_group(self, start): def _find_group(self, current_board, vertex):
color = self.simulate_board[self.game._flatten(start)] color = current_board[self.game._flatten(vertex)]
# print ("color : ", color) # print ("color : ", color)
chain = set() chain = set()
frontier = [start] frontier = [vertex]
has_liberty = False has_liberty = False
while frontier: while frontier:
current = frontier.pop() current = frontier.pop()
# print ("current : ", current) # print ("current : ", current)
chain.add(current) chain.add(current)
for n in self._neighbor(current): for n in self._neighbor(current):
# print n, self._flatten(n), self.board[self._flatten(n)], if current_board[self.game._flatten(n)] == color and not n in chain:
if self.simulate_board[self.game._flatten(n)] == color and not n in chain:
frontier.append(n) frontier.append(n)
if self.simulate_board[self.game._flatten(n)] == utils.EMPTY: if current_board[self.game._flatten(n)] == utils.EMPTY:
has_liberty = True has_liberty = True
return has_liberty, chain return has_liberty, chain
def _is_suicide(self, color, vertex): def _is_suicide(self, current_board, color, vertex):
self.simulate_board[self.game._flatten(vertex)] = color # assume that we already take this move current_board[self.game._flatten(vertex)] = color # assume that we already take this move
suicide = False suicide = False
has_liberty, group = self._find_group(vertex) has_liberty, group = self._find_group(current_board, vertex)
if not has_liberty: if not has_liberty:
suicide = True # no liberty, suicide suicide = True # no liberty, suicide
for n in self._neighbor(vertex): for n in self._neighbor(vertex):
if self.simulate_board[self.game._flatten(n)] == utils.another_color(color): if current_board[self.game._flatten(n)] == utils.another_color(color):
opponent_liberty, group = self._find_group(n) opponent_liberty, group = self._find_group(current_board, n)
if not opponent_liberty: if not opponent_liberty:
suicide = False # this move is able to take opponent's stone, not suicide suicide = False # this move is able to take opponent's stone, not suicide
self.simulate_board[self.game._flatten(vertex)] = utils.EMPTY # undo this move current_board[self.game._flatten(vertex)] = utils.EMPTY # undo this move
return suicide return suicide
def _check_global_isomorphous(self, color, vertex): def _check_global_isomorphous(self, history_boards, current_board, color, vertex):
##backup repeat = False
_board = copy.copy(self.simulate_board) next_board = copy.copy(current_board)
self.simulate_board[self.game._flatten(vertex)] = color next_board[self.game._flatten(vertex)] = color
self._process_board(color, vertex) self._process_board(next_board, color, vertex)
if self.simulate_board in self.game.history: if next_board in history_boards:
res = True repeat = True
else: return repeat
res = False
self.simulate_board = _board
return res
def _in_board(self, vertex): def _in_board(self, vertex):
x, y = vertex x, y = vertex
@ -92,28 +87,28 @@ class GoEnv:
corner.append((_x, _y)) corner.append((_x, _y))
return corner return corner
def _process_board(self, color, vertex): def _process_board(self, current_board, color, vertex):
nei = self._neighbor(vertex) nei = self._neighbor(vertex)
for n in nei: for n in nei:
if self.simulate_board[self.game._flatten(n)] == utils.another_color(color): if current_board[self.game._flatten(n)] == utils.another_color(color):
has_liberty, group = self._find_group(n) has_liberty, group = self._find_group(current_board, n)
if not has_liberty: if not has_liberty:
for b in group: for b in group:
self.simulate_board[self.game._flatten(b)] = utils.EMPTY current_board[self.game._flatten(b)] = utils.EMPTY
def _is_eye(self, color, vertex): def _is_eye(self, current_board, color, vertex):
nei = self._neighbor(vertex) nei = self._neighbor(vertex)
cor = self._corner(vertex) cor = self._corner(vertex)
ncolor = {color == self.simulate_board[self.game._flatten(n)] for n in nei} ncolor = {color == current_board[self.game._flatten(n)] for n in nei}
if False in ncolor: if False in ncolor:
# print "not all neighbors are in same color with us" # print "not all neighbors are in same color with us"
return False return False
_, group = self._find_group(nei[0]) _, group = self._find_group(current_board, nei[0])
if set(nei) < group: if set(nei) < group:
# print "all neighbors are in same group and same color with us" # print "all neighbors are in same group and same color with us"
return True return True
else: else:
opponent_number = [self.simulate_board[self.game._flatten(c)] for c in cor].count(-color) opponent_number = [current_board[self.game._flatten(c)] for c in cor].count(-color)
opponent_propotion = float(opponent_number) / float(len(cor)) opponent_propotion = float(opponent_number) / float(len(cor))
if opponent_propotion < 0.5: if opponent_propotion < 0.5:
# print "few opponents, real eye" # print "few opponents, real eye"
@ -122,49 +117,54 @@ class GoEnv:
# print "many opponents, fake eye" # print "many opponents, fake eye"
return False return False
def knowledge_prunning(self, color, vertex): def knowledge_prunning(self, current_board, color, vertex):
### check if it is an eye of yourself ### check if it is an eye of yourself
### assumptions : notice that this judgement requires that the state is an endgame ### assumptions : notice that this judgement requires that the state is an endgame
if self._is_eye(color, vertex): if self._is_eye(current_board, color, vertex):
return False return False
return True return True
def simulate_is_valid(self, state, action): def sa2cv(self, state, action):
# State is the play board, the shape is [1, self.game.size, self.game.size, 17]. # State is the play board, the shape is [1, self.game.size, self.game.size, 17], action is an index.
# Action is an index
# We need to transfer the (state, action) pair into (color, vertex) pair to simulate the move # We need to transfer the (state, action) pair into (color, vertex) pair to simulate the move
if action == self.game.size ** 2:
vertex = (0, 0)
else:
vertex = self.game._deflatten(action)
if state[0, 0, 0, -1] == utils.BLACK: if state[0, 0, 0, -1] == utils.BLACK:
color = utils.BLACK color = utils.BLACK
else: else:
color = utils.WHITE color = utils.WHITE
if action == self.game.size ** 2:
vertex = (0, 0)
else:
vertex = self.game._deflatten(action)
return color, vertex
def simulate_is_valid(self, history_boards, current_board, state, action):
# initialize simulate_latest_boards and simulate_board from state
self.simulate_latest_boards.clear() self.simulate_latest_boards.clear()
for i in range(8): for i in range(8):
self.simulate_latest_boards.append((state[:, :, :, i] - state[:, :, :, i + 8]).reshape(-1).tolist()) self.simulate_latest_boards.append((state[:, :, :, i] - state[:, :, :, i + 8]).reshape(-1).tolist())
self.simulate_board = copy.copy(self.simulate_latest_boards[-1]) self.simulate_board = copy.copy(self.simulate_latest_boards[-1])
color, vertex = self.sa2cv(state, action)
### in board ### in board
if not self._in_board(vertex): if not self._in_board(vertex):
return False return False
### already have stone ### already have stone
if not self.simulate_board[self.game._flatten(vertex)] == utils.EMPTY: if not current_board[self.game._flatten(vertex)] == utils.EMPTY:
# print(np.array(self.board).reshape(9, 9)) # print(np.array(self.board).reshape(9, 9))
# print(vertex) # print(vertex)
return False return False
### check if it is suicide ### check if it is suicide
if self._is_suicide(color, vertex): if self._is_suicide(current_board, color, vertex):
return False return False
### forbid global isomorphous ### forbid global isomorphous
if self._check_global_isomorphous(color, vertex): if self._check_global_isomorphous(history_boards, current_board, color, vertex):
return False return False
if not self.knowledge_prunning(color, vertex): if not self.knowledge_prunning(current_board, color, vertex):
return False return False
return True return True
@ -181,17 +181,11 @@ class GoEnv:
return False return False
def simulate_step_forward(self, state, action): def simulate_step_forward(self, state, action):
if state[0, 0, 0, -1] == 1: # initialize the simulate_board from state
color = utils.BLACK
else:
color = utils.WHITE
if action == self.game.size ** 2:
vertex = utils.PASS
else:
vertex = self.game._deflatten(action)
# print(vertex)
# print(self.board)
self.simulate_board = (state[:, :, :, 7] - state[:, :, :, 15]).reshape(-1).tolist() self.simulate_board = (state[:, :, :, 7] - state[:, :, :, 15]).reshape(-1).tolist()
color, vertex = self.sa2cv(state, action)
self.simulate_do_move(color, vertex) self.simulate_do_move(color, vertex)
new_state = np.concatenate( new_state = np.concatenate(
[state[:, :, :, 1:8], (np.array(self.simulate_board) == utils.BLACK).reshape(1, self.game.size, self.game.size, 1), [state[:, :, :, 1:8], (np.array(self.simulate_board) == utils.BLACK).reshape(1, self.game.size, self.game.size, 1),

3
tianshou/core/mcts/mcts.py
View File

@ -75,7 +75,8 @@ class UCTNode(MCTSNode):
start_time = time.time() start_time = time.time()
self.mask = [] self.mask = []
for act in range(self.action_num - 1): for act in range(self.action_num - 1):
if not simulator.simulate_is_valid(self.state, act): if not simulator.simulate_is_valid(
simulator.simulate_latest_boards, simulator.simulate_board, self.state, act):
self.mask.append(act) self.mask.append(act)
self.ucb[act] = -float("Inf") self.ucb[act] = -float("Inf")
else: else: