def validate_move(
board_list: List[List[Piece]],
from_row: int,
from_col: int,
to_row: int,
to_col: int,
) -> bool:
"""Validate move given by current and desired piece coordinates.
Args:
board_list: Information about positions of pieces.
from_row: Row of board of original piece location.
from_col: Column of board of original piece location.
to_row: Cow of board of desired piece location.
to_col: Column of board of desired piece location.
Returns:
True if given move is valid, otherwise false.
"""
# not among available moves
if (to_row,
to_col) not in Rules.generate_all_moves(from_row, from_col):
return False
# can't move piece from outside of board
if from_row < 0 or from_col < 0 or from_row > 7 or from_col > 7:
return False
# cant move out of board
if to_row < 0 or to_col < 0 or to_row > 7 or to_col > 7:
return False
# target square must be empty
if board_list[to_row][to_col] is not None:
return False
# can't move empty square
p = board_list[from_row][from_col]
if p is None:
return False
# cant move in opposite direction, except king
if p.direction == Constants().UP and from_row < to_row and not p.king:
return False
if p.direction == Constants(
).DOWN and from_row > to_row and not p.king:
return False
# cant jump over itself or empty square
between_row, between_col = Rules.get_between_position(
from_row, from_col, to_row, to_col)
if between_row is not None and between_col is not None:
pp = board_list[between_row][between_col]
if pp is None or pp.ptype == p.ptype:
return False
return True
def __init__(self,
name: str,
ptype: int,
load_model: bool = True,
epsilon: float = 0.0):
self.board_enc = BoardEncoding()
if ptype == Constants().DARK:
self.board_enc.dark = 0.5
self.board_enc.light = -0.5
self.board_enc.dark_king = 1.
self.board_enc.light_king = -1.
elif ptype == Constants().LIGHT:
self.board_enc.dark = -0.5
self.board_enc.light = 0.5
self.board_enc.dark_king = -1.
self.board_enc.light_king = 1.
else:
raise ValueError
super().__init__(name, ptype)
self.render = False
self.load_model = load_model
# 상태와 행동의 크기 정의
#self.state_size = (8, 8, 1)
self.action_size = 4
# DQN 하이퍼파라미터
self.discount_factor = 0.99
self.learning_rate = 0.001
self.epsilon = epsilon
self.epsilon_decay = 0.9999
self.epsilon_min = 0.00
self.batch_size = 64
self.train_start = 3000
# 리플레이 메모리, 최대 크기 2000
self.memory = deque(maxlen=4000)
# 모델과 타깃 모델 생성
self.model = self.build_model()
self.target_model = self.build_model()
# 타깃 모델 초기화
self.update_target_model()
if self.load_model:
self.model.load_weights("./save_model/checker_dqn.h5")
def test_become_king(self, empty_board):
# Dark piece
empty_board.board_list[5][5] = DarkPiece()
empty_board.move(Constants().DARK, 5, 5, 6, 6)
assert empty_board.board_list[6][6].is_king() is False
empty_board.move(Constants().DARK, 6, 6, 7, 7)
assert empty_board.board_list[7][7].is_king() is True
# Light piece
empty_board.board_list[2][2] = LightPiece()
empty_board.move(Constants().LIGHT, 2, 2, 1, 1)
assert empty_board.board_list[1][1].is_king() is False
empty_board.move(Constants().LIGHT, 1, 1, 0, 0)
assert empty_board.board_list[0][0].is_king() is True
def get_opponent_type(ptype: int, ) -> int:
"""Get a type of opponent agent.
Note: In checkers there is only one pair of agents competing with each other.
Args:
ptype: Type of piece.
Returns:
opponent_type: Type of opponent agent.
"""
if ptype == Constants().DARK:
opponent_type = Constants().LIGHT
else:
opponent_type = Constants().DARK
return opponent_type
def __init__(
self,
ptype: int,
):
"""Initialize random agent.
Args:
name: name of agent.
ptype: type of piece that agent is responsible for.
"""
if ptype == Constants().DARK:
name = "HumanDark"
elif ptype == Constants().LIGHT:
name = "HumanLight"
else:
raise ValueError
super().__init__(name, ptype)
def __init__(self):
self._constants = Constants()
self._encoding = {}
self.empty = 0
self.dark = 10
self.dark_king = 11
self.light = 20
self.light_king = 21
def board_list2numpy(
board_list: List[List],
) -> np.array:
board_size = len(board_list)
board_numpy = Constants().EMPTY * np.ones((board_size, board_size))
for row in range(board_size):
for col in range(board_size):
if board_list[row][col] is not None:
board_numpy[row][col] = board_list[row][col].ptype
return board_numpy
def main():
max_episodes = 10000
with tf.Session() as sess:
# saver = tf.train.Saver()
####### Agent Setting #######
MasterAgent = RandomAgentLight("Teacher Agent")
MyAgent = DqnAgent(sess, "doublejtoh Agent", Constants().LIGHT)
tf.global_variables_initializer().run()
current_agent = MasterAgent
next_agent = MyAgent
for episode in range(max_episodes):
step = 0
done = False
obs = env.reset()
env.render()
while not done:
state = board_list2numpy(obs)
if current_agent._name == 'doublejtoh Agent':
from_row, from_col, to_row, to_col = current_agent.act(state, episode)
# print(from_row, from_col, to_row, to_col)
else:
from_row, from_col, to_row, to_col = current_agent.act(obs)
# print("Teacher Agent: ", from_row, from_col, to_row, to_col)
obs, reward, done, info = env.step(current_agent, from_row, from_col, to_row, to_col)
# print(info)
if done:
print(f"{current_agent} agent wins.")
action = (to_row, to_col)
next_state = board_list2numpy(obs)
if current_agent._name == 'doublejtoh Agent':
current_agent.consume(state, action, reward, done, episode, next_state) # here, obs means new state. save to memory.
### change turn ###
temporary_agent = current_agent
current_agent = next_agent
next_agent = temporary_agent
env.render()
step += 1
print("Episode ", episode, "step: ", step)
# if current_agent._name == 'doublejtoh Agent':
# current_agent.consume_after_episode(episode) # replay train.
MyAgent.consume_after_episode(episode)
def board_list2numpy(
board_list: List[List],
encoding: BoardEncoding=BoardEncoding(),
) -> np.array:
"""Convert the state of game (`board_list`) into 2D NumPy Array using `encoding`.
Args:
board_list: (List[List[Piece]]) State of the game.
encoding: (BoardEncoding) Optional argument. If not given default encoding will be utilized.
Returns:
board_numpy: (np.array)
"""
board_size = len(board_list)
constants = Constants()
board_numpy = encoding[constants.EMPTY] * np.ones((board_size, board_size))
for row in range(board_size):
for col in range(board_size):
if board_list[row][col] is not None:
ptype = board_list[row][col].ptype
king = board_list[row][col].king
if ptype == constants.LIGHT:
if king:
piece_type = constants.LIGHT_KING
else:
piece_type = constants.LIGHT
else: # DARK
if king:
piece_type = constants.DARK_KING
else:
piece_type = constants.DARK
board_numpy[row][col] = encoding[piece_type]
return board_numpy
def test_move(self, empty_board):
# Dark piece
empty_board.board_list[5][5] = DarkPiece()
with pytest.raises(ValueError):
empty_board.move(Constants().DARK, 5, 5, 4, 4)
# move any direction
empty_board.board_list[5][5].make_king()
empty_board.move(Constants().DARK, 5, 5, 4, 4)
empty_board.move(Constants().DARK, 4, 4, 5, 5)
# Light Piece
empty_board.board_list[5][5] = LightPiece()
with pytest.raises(ValueError):
empty_board.move(Constants().LIGHT, 5, 5, 6, 6)
# move any direction
empty_board.board_list[5][5].make_king()
empty_board.move(Constants().LIGHT, 5, 5, 6, 6)
empty_board.move(Constants().LIGHT, 6, 6, 5, 5)
def test_move(self, empty_board):
# Dark piece
init_board = empty_board
empty_board.board_list[5][5] = DarkPiece()
empty_board.move(Constants().DARK, 5, 5, 4, 4)
assert init_board == empty_board
# move any direction
empty_board.board_list[5][5].make_king()
empty_board.move(Constants().DARK, 5, 5, 4, 4)
empty_board.move(Constants().DARK, 4, 4, 5, 5)
# Light Piece
init_board = empty_board
empty_board.board_list[5][5] = LightPiece()
empty_board.move(Constants().LIGHT, 5, 5, 6, 6)
assert init_board == empty_board
# move any direction
empty_board.board_list[5][5].make_king()
empty_board.move(Constants().LIGHT, 5, 5, 6, 6)
empty_board.move(Constants().LIGHT, 6, 6, 5, 5)
def __init__(
self,
name: str,
):
super().__init__(name, Constants().LIGHT)
def light():
return Constants().LIGHT
def dark():
return Constants().DARK
from agent import DQNChecker
if __name__ == "__main__":
args = argparse.ArgumentParser()
args.add_argument('--render', type=bool, default=False)
args.add_argument('--episodes', type=int, default=3000)
args.add_argument('--train', type=bool, default=False)
args.add_argument('--thresold', type=int, default=3000)
config = args.parse_args()
env = gym.make("Checkers")
if config.train:
a1 = DQNChecker("Agent_1", Constants().DARK, True, 0.1)
a2 = DQNChecker("Agent_2", Constants().LIGHT, True, 0.1)
else:
a1 = DQNChecker("Agent_1", Constants().DARK, True, 0.0)
a2 = DQNChecker("Agent_2", Constants().LIGHT, True, 0.0)
history = {}
history[a1] = {'scores': [], 'episodes': []}
history[a2] = {'scores': [], 'episodes': []}
agent_tag = {}
agent_tag[a1] = 'Agent_1'
agent_tag[a2] = 'Agent_2'
for e in range(config.episodes):
done = False
score = 0
def __init__(self, ):
super().__init__(Constants().DARK)
def __init__(self, ):
super().__init__(Constants().LIGHT)
