def get_move(self, board: Board) -> int:
"""
Return the next move given the board `board` based on the current values of next states
:param board: The current board state
:return: The next move based on the current values of next states, starting from input state
"""
if self.move_strategy == MoveStrategy.EXPLORATION:
# exploratory random move
m = board.random_empty_spot()
_ = self.get_v(
board) # just to ensure we have values for our board state
return m
else:
# greedy move: exploiting current knowledge
vals = self.get_v(board) # type: np.ndarray
while True:
maxv_idxs = np.argwhere(
vals == np.amax(vals)) # positions of max values in array
m = np.random.choice(maxv_idxs.flatten().tolist()) # type: int
#m = np.argmax(vals) # type: int # this instead would return 1st occurance
if board.is_legal(m):
# print("vals=", end='')
# print(vals)
# print("m={}".format(m))
return m
else:
vals[m] = -1.0
def move(self, board: Board) -> (GameResult, bool):
"""
Makes a move on the given input state
:param board: The current state of the game
:return: The GameResult after this move, Flag to indicate whether the move finished the game
"""
self.board_position_log.append(board.state.copy())
nn_input = self.board_state_to_nn_input(board.state)
probs = self.get_valid_probs([nn_input], [board])
probs = probs[0]
# Most of the time our next move is the one with the highest probability after removing all illegal ones.
# Occasionally, however we randomly chose a random move to encourage exploration
if (self.training is True) and \
(self.game_counter < self.pre_training_games):
move = board.random_empty_spot()
else:
if np.isnan(probs).any(): # Can happen when all probabilities degenerate to 0. Best thing we can do is
# make a random legal move
move = board.random_empty_spot()
else:
move = np.random.choice(np.arange(len(probs)), p=probs)
if not board.is_legal(move): # Debug case only, I hope
print("Illegal move!")
# We record the action we selected as well as the Q values of the current state for later use when
# adjusting NN weights.
self.action_log.append(move)
_, res, finished = board.move(move, self.side)
return res, finished
def get_v(self, board: Board) -> np.ndarray:
"""
Returns all values when moving from current state of 'board'
:param board: The current board state
:return: List of values of all possible next board states
"""
# We build the value dictionary in a lazy manner, only adding a state when it is actually used for the first time
#
board_hash = board.hash_value(
) # needed because value dictionary maps *hashed* state to values
if board_hash in self.v:
vals = self.v[board_hash]
else:
vals = np.full(9, self.v_init) # default initial value
# set values for winning states to WIN_VALUE
# (player cannot end up in a losing state after a move
# so losing states need not be considered):
for pos in range(vals.size): # vals.size = BOARD_SIZE
if board.is_legal(pos):
b = Board(board.state)
b.move(pos, self.side)
if b.check_win():
vals[pos] = self.v_win
elif b.num_empty() == 0:
# if it is not a win, and there are no other positions
# available, then it is a draw
vals[pos] = self.v_draw
# Update dictionary:
self.v[board_hash] = vals
# print("v[{}]={}".format(board_hash, self.v[board_hash]))
return vals
def get_move(self, board: Board) -> int:
"""
Return the next move given the board `board` based on the current Q values
:param board: The current board state
:return: The next move based on the current Q values for the input state
"""
board_hash = board.hash_value() # type: int
qvals = self.get_q(board_hash) # type: np.ndarray
while True:
m = np.argmax(qvals) # type: int
if board.is_legal(m):
return m
else:
qvals[m] = -1.0
def move(self, board: Board) -> (GameResult, bool):
"""
Implements the Player interface and makes a move on Board `board`
:param board: The Board to make a move on
:return: A tuple of the GameResult and a flag indicating if the game is over after this move.
"""
# We record all game positions to feed them into the NN for training with the corresponding updated Q
# values.
self.board_position_log.append(board.state.copy())
nn_input = self.board_state_to_nn_input(board.state)
probs, qvalues = self.get_probs(nn_input)
qvalues = np.copy(qvalues)
# We filter out all illegal moves by setting the probability to 0. We don't change the q values
# as we don't want the NN to waste any effort of learning different Q values for moves that are illegal
# anyway.
for index, p in enumerate(qvalues):
if not board.is_legal(index):
probs[index] = -1
elif probs[index] < 0:
probs[index] = 0.0
# Most of the time our next move is the one with the highest probability after removing all illegal ones.
# Occasionally, however we randomly chose a random move to encourage exploration
if (self.training is True) and (np.random.rand(1) <
self.random_move_prob):
move = board.random_empty_spot()
else:
move = np.argmax(probs)
# Unless this is the very first move, the max Q value of this state is also the max Q value of
# the move that got the game from the previous state to this one.
if len(self.action_log) > 0:
self.next_max_log.append(qvalues[np.argmax(probs)])
# We record the action we selected as well as the Q values of the current state for later use when
# adjusting NN weights.
self.action_log.append(move)
self.values_log.append(qvalues)
# We execute the move and return the result
_, res, finished = board.move(move, self.side)
return res, finished
