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 _min(self, board: Board) -> (float, int):
"""
Evaluate the board position `board` from the Minimizing player's point of view.
:param board: The board position to evaluate
:return: Tuple of (Best Result, Best Move in this situation). Returns -1 for best move if the game has already
finished
"""
#
# First we check if we have seen this board position before, and if yes just return the cached value
#
board_hash = board.hash_value()
if board_hash in self.cache:
return self.cache[board_hash]
#
# Init the min value as well as action. Min value is set to DRAW as this value will pass through in case
# of a draw
#
min_value = self.DRAW_VALUE
action = -1
#
# If the game has already finished we return. Otherwise we look at possible continuations
#
winner = board.who_won()
if winner == self.side:
min_value = self.WIN_VALUE
action = -1
elif winner == board.other_side(self.side):
min_value = self.LOSS_VALUE
action = -1
else:
for index in [
i for i, e in enumerate(board.state)
if board.state[i] == EMPTY
]:
b = Board(board.state)
b.move(index, board.other_side(self.side))
res, _ = self._max(b)
if res < min_value or action == -1:
min_value = res
action = index
# Shortcut: Can't get better than that, so abort here and return this move
if min_value == self.LOSS_VALUE:
self.cache[board_hash] = (min_value, action)
return min_value, action
self.cache[board_hash] = (min_value, action)
return min_value, action
def _min(self, board: Board) -> int:
"""
Evaluate the board position `board` from the Minimizing player's point of view.
:param board: The board position to evaluate
:return: returns the best Move in this situation. Returns -1 for best move if the game has already
finished
"""
#
# First we check if we have seen this board position before, and if yes just return a random choice
# from the cached values
#
board_hash = board.hash_value()
if board_hash in self.cache:
return random.choice(self.cache[board_hash])
#
# If the game has already finished we return. Otherwise we look at possible continuations
#
winner = board.who_won()
if winner == self.side:
best_moves = {(self.WIN_VALUE, -1)}
elif winner == board.other_side(self.side):
best_moves = {(self.LOSS_VALUE, -1)}
else:
#
# Init the min value as well as action. Min value is set to DRAW as this value will pass through in case
# of a draw
#
min_value = self.DRAW_VALUE
action = -1
best_moves = {(min_value, action)}
for index in [
i for i, e in enumerate(board.state)
if board.state[i] == EMPTY
]:
b = Board(board.state)
b.move(index, board.other_side(self.side))
res, _ = self._max(b)
if res < min_value or action == -1:
min_value = res
action = index
best_moves = {(min_value, action)}
elif res == min_value:
action = index
best_moves.add((min_value, action))
best_moves = tuple(best_moves)
self.cache[board_hash] = best_moves
return random.choice(best_moves)
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, _ = self.get_valid_probs([nn_input], self.q_net, [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) or (np.random.rand(1) < self.random_move_prob)):
move = board.random_empty_spot()
else:
move = 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)
# We execute the move and return the result
_, res, finished = board.move(move, self.side)
return res, finished
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 move(self, board: Board) -> (GameResult, bool):
"""
Making a random move
:param board: The board to make a move on
:return: The result of the move
"""
_, res, finished = board.move(board.random_empty_spot(), self.side)
return res, finished
def move(self, board: Board) -> (GameResult, bool):
"""
Making a move according to the MinMax algorithm
:param board: The board to make a move on
:return: The result of the move
"""
score, action = self._max(board)
_, res, finished = board.move(action, self.side)
return res, finished
def move(self, board: Board):
"""
Makes a move and returns the game result after this move and whether the move ended the game
:param board: The board to make a move on
:return: The GameResult after this move, Flag to indicate whether the move finished the game
"""
m = self.get_move(board)
self.move_history.append((board.hash_value(), m))
_, res, finished = board.move(m, self.side)
return res, finished
def move(self, board: Board) -> (GameResult, bool):
"""
Making a move according to the MinMax algorithm. If more than one best move exist, chooses amongst them
randomly.
:param board: The board to make a move on
:return: The result of the move
"""
score, action = self._max(board)
_, res, finished = board.move(action, self.side)
return res, finished
def play_random_game():
board = Board()
finished = False
last_play = NAUGHT
next_play = CROSS
while not finished:
_, result, finished = board.move(board.random_empty_spot(), next_play)
print_board(board)
last_play, next_play = next_play, last_play
if result == GameResult.DRAW:
print("Game is a draw")
elif last_play == CROSS:
print("Cross won!")
else:
print("Naught won!")
def move(self, board: Board) -> (GameResult, bool):
"""
Make move corresponding to key pressed by user
:param board: The board to make a move on
:return: The result of the move
"""
print()
while True:
key = input("Your move? ")
if key in self.keys:
break
position = self.keys.index(key)
_, res, finished = board.move(position, self.side)
return res, finished
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
def move(self, board: Board):
"""
Makes a move and returns the game result after this move and whether the move ended the game
:param board: The board to make a move on
:return: The GameResult after this move, Flag to indicate whether the move finished the game
"""
# Select strategy to choose next move: exploit known or explore unknown?
if np.random.uniform(0, 1) <= self.epsilon:
self.move_strategy = MoveStrategy.EXPLORATION
else:
self.move_strategy = MoveStrategy.EXPLOITATION
m = self.get_move(board)
self.move_history.append((board.hash_value(), m))
self.backup_value()
# print("v={}".format(self.v))
_, res, finished = board.move(m, self.side)
return res, finished
