def test_if_three_in_a_row_wins(self):
board = Board()
tile_type = 1 # White
board.add_tile(tile_type, row=1, column=1)
board.add_tile(tile_type, row=1, column=0)
board.add_tile(tile_type, row=1, column=2)
self.assertEqual(board.three_in_a_row, True)
def test_a_cell_with_2or3_neighbours_stays_alive_after_tick(self):
my_board = Board([self.fst_cell, self.snd_cell, self.trd_cell, self.fourth_cell])
my_board.tick()
self.assertTrue(my_board.cell_alive_at(6, 6))
self.assertTrue(my_board.cell_alive_at(5, 6))
self.assertTrue(my_board.cell_alive_at(5, 7))
self.assertTrue(my_board.cell_alive_at(6, 7))
def test_combine(self):
board = Board(4, 4)
start_single = [2, 2, 0, 0]
start_double = [2, 2, 4, 4]
end_single = [4, 0, 0, 0]
end_double = [4, 0, 8, 0]
self.assertEqual(end_single, board._Board__combine(start_single))
self.assertEqual(end_double, board._Board__combine(start_double))
def ai_move():
layout = request.args.get('layout')
layout = json.loads(layout)
board = Board(setup=layout)
ai.move(board)
layout = board.board()
layout = json.dumps(layout)
return layout
def test_transpose(self):
board = Board()
tile_type = 1 # White
board.add_tile(tile_type, row=1, column=1)
board.add_tile(tile_type, row=1, column=0)
board.add_tile(tile_type, row=1, column=2)
#board.transpose()
self.assertEqual(str(board), "")
def test_compress(self):
board = Board(4, 4)
start = [2, 0, 4, 0]
end_backward = [2, 4, 0, 0]
end_forward = [0, 0, 2, 4]
self.assertEqual(end_backward, board._Board__compress(start, 'left'))
self.assertEqual(end_forward, board._Board__compress(start, 'right'))
self.assertEqual(end_backward, board._Board__compress(start, 'up'))
self.assertEqual(end_forward, board._Board__compress(start, 'down'))
def test_set(self):
board = Board(4, 4)
end = [
1, 0, 0, 0,
0, 0, 0, 0,
0, 0, 0, 0,
0, 0, 0, 0
]
board._Board__set(0, 0, 1)
self.assertEqual(end, board.grid)
class TestBoard(unittest.TestCase):
def setUp(self):
self.board = Board(10,10)
def test_initialise_collection(self):
self.board.initialise_collection()
self.assertEqual(len(self.board.collection), 100)
def test_get_neighbours(self):
for i in range (100):
self.board.collection.append(Cell(False))
def test_board_join(self):
"""
When player2 join to the game, the pieces are in the board.
"""
player1 = MockPlayer()
player2 = MockPlayer()
board1 = Board(player1)
board1.join(player2)
self.assertEqual(board1._pieces[(0, 0)], Piece(player1, 7))
class StackingTests(unittest.TestCase):
def setUp(self):
self.board = Board()
def test_stack_one_red_on_one_blue(self):
red = self.board.get_colour('red')
blue = self.board.get_colour('blue')
new_red, new_blue = stack(red, blue)
self.assertEqual(red.top, new_blue.top)
self.assertEqual([], new_red)
self.assertEqual(2, new_blue.height)
def setup_board(players, num_of_nodes):
Setup().setup_logger()
nodes = {}
board = Board(nodes)
#create empty nodes
for i in range(0, num_of_nodes):
nodes[i] = board.Node(i, 0, 1)
#create connections
for i in range(1, num_of_nodes):
board.connect_nodes(i, i - 1)
board.connect_nodes(0, num_of_nodes - 1)
for i in range(0, num_of_nodes/4):
board.connect_nodes(randint(0, num_of_nodes - 1), randint(0, num_of_nodes - 1))
#split nodes between players
nodes_to_split = nodes.values()[:]
shuffle(nodes_to_split)
for i in range(0, num_of_nodes):
player_id = i % len(players)
nodes_to_split[i].owner = player_id
#split armies between players
armies_per_node = 3
armies_per_player = num_of_nodes * (armies_per_node - 1) / len(players)
for player in players:
my_nodes = board.nodes_by_owner(player.id)
for i in range(0, armies_per_player):
node = my_nodes[randint(0, len(my_nodes) - 1)]
node.add_army(1)
return board
def test_new_tile(self):
start = [
0, 0, 0, 0,
0, 0, 0, 0,
0, 0, 0, 0,
0, 0, 0, 0
]
board = Board(4, 4)
self.assertEqual(start, board.grid)
board.new_tile()
self.assertNotEqual(start, board.grid)
def test_traverse(self):
board = Board()
paths = board.traverse()
# horizontals
for y in range(0, 3):
self.assertIn([(0, y), (1, y), (2, y)], paths)
# verticals
for x in range(0, 3):
self.assertIn([(x, 0), (x, 1), (x, 2)], paths)
# \ diagonal
self.assertIn([(0, 0), (1, 1), (2, 2)], paths)
# / diagonal
self.assertIn([(0, 2), (1, 1), (2, 0)], paths)
def setUp(self):
player1 = MockPlayer()
player2 = MockPlayer()
board1 = Board(player1)
board1.join(player2)
board1._pieces = {
(0, 0): Piece(player1, 7),
(8, 8): Piece(player2, 7),
(0, 3): Hole(player1),
(8, 3): Hole(player2),
}
self.board = board1
self.player1 = player1
self.player2 = player2
def test_double_combine(self):
start = [
1, 1, 1, 1,
1, 1, 2, 2,
2, 2, 1, 1,
0, 0, 0, 0
]
end = [
2, 2, 0, 0,
2, 3, 0, 0,
3, 2, 0, 0,
0, 0, 0, 0
]
board = Board(4, 4, start)
board.move('left')
self.assertEqual(end, board.grid)
def test_combine_down(self):
start = [
0, 0, 1, 1,
0, 1, 0, 1,
1, 0, 0, 0,
1, 1, 1, 0
]
end = [
0, 0, 0, 0,
0, 0, 0, 0,
0, 0, 0, 0,
2, 2, 2, 2
]
board = Board(4, 4, start)
board.move('down')
self.assertEqual(end, board.grid)
def test_combine_left(self):
start = [
1, 1, 0, 0,
1, 0, 1, 0,
1, 0, 0, 1,
0, 0, 1, 1
]
end = [
2, 0, 0, 0,
2, 0, 0, 0,
2, 0, 0, 0,
2, 0, 0, 0
]
board = Board(4, 4, start)
board.move('left')
self.assertEqual(end, board.grid)
def __init__(self, w, h, owner):
self.sid = sha1(str(time.time())).hexdigest()
self.clients = [owner]
owner.player = 1
self.player_counter = 2
self.state = STATE_JOINING
self.board = Board(w, h)
self.turn = owner
owner.send(Msg('newgame', self.sid, w, h).encode())
owner.send(Msg('turn', self.turn.player).encode())
def main():
solver = Solver()
manual = Player()
n = 17
if(len(sys.argv)>1):
n=int(sys.argv[1]);
# Start simulator on Circle with 10 points with solver as player 0
# manual (you) as player 1.
sim = Simulator(Board.from_circle(n),solver,manual)
print("Starting simulation on circle with <0> points. The computer is\
playing first!\n")
while(sim.step()): pass;
input()
def start(self, PLAYERS):
self.board = Board.initial(GUICard)
# Clear the ingame data.
self.start_turn_time = time.clock()
self.set_time = 0
self.players = {x: [] for x in xrange(PLAYERS)}
#self.cards_drawn = False
self.table_clickable = False # Is it possible to select cards?
self.draw_END_BUTTON = False # Draw 'No Sets' and block the game?
self.player_turn = None # Whose turn (== player_id or None)
def server():
import cv2
s = socket.socket()
host = socket.gethostname()
print(host)
print(socket.gethostbyname(socket.gethostname()))
host = HOST
port = PORT
s.bind((host, port))
s.listen(5)
while True:
print('listening...')
game = Game(handicap=HANDICAP)
board_img = game.get_current_board_img()
cv2.imshow('board_img', board_img)
param = {'MCTS': False}
cv2.setMouseCallback('board_img', game.cap_click, param=param)
cv2.waitKey(33)
c, addr = s.accept()
print('Got connection from', addr)
while True:
before_len = len(game.boards)
board_img = game.get_current_board_img(last_move=game.current_moves[-1])
cv2.imshow('board_img', board_img)
cv2.waitKey(33)
now_len = len(game.boards)
if now_len > before_len:
print(param['MCTS'])
board_img = game.get_current_board_img(last_move=game.current_moves[-1])
cv2.imshow('board_img', board_img)
cv2.waitKey(33)
latest_board = game.boards[-2] # board before human move
next_to_play = game.next_to_play
board_str = Board.mtx2str(latest_board.board_mtx)
next_to_play = str(next_to_play)
print('next_to_play:', next_to_play)
c.send(str.encode(str(game.current_moves[-1]) + '|' + next_to_play + '|' + board_str + '|'
+ str(int(param['MCTS']))))
print(str(game.current_moves[-1]))
move = c.recv(1024).decode('utf-8')
print('move', move)
temp = move.split(',')
x, y = int(temp[0][1:]), int(temp[1][1:-1])
print(x, y, game.next_to_play)
game.mk_move(x, y)
c.close()
def main():
p1 = Player(name="John")
p2 = Player(name="Hulio")
b = Board()
b.add_player(p1)
b.add_player(p2)
b.start()
_pprint(filled_cells(b.state))
raw_input()
destruct_status = False
prev_check = []
state = b.state
promotions = 0
while state not in ("FIN", ):
# for player in (p1, p2):
# b.get_state(player)
event = smart_event_factory(state, colors, destruct_status, prev_check)
color = 'white' if 'w' == event["name"][0] else 'black'
if check_the_turn(event, state) == False:
continue
else:
print "VALID"
if "promoted" in event.keys() and event["promoted"]:
promotions +=1
print "PROMOTED CHECKS: ", promotions
state = check_the_game_state(state, event, colors)
if event['cutDown'] and destruction_is_possible(event['finalPos'], filled_cells(state, color=colors[1]), filled_cells(state, color=colors[0]), event.get('promoted')):
destruct_status = True
prev_check = event['finalPos']
print 'NEXT DESTRUCT'
raw_input()
else:
colors.reverse()
destruct_status = False
prev_check = []
print 'COLORS REVERSED!'
print 'FILLED CELLS: ' , len((filled_cells(state)))
raw_input()
def __init__(self, width, height, players_num, left_player=Soldier, right_player=Soldier):
"""Initializes game board with width and height,
adds players_num playes on both sides
addional parameters left_player and right_player can be set
to different subclasses of Player's
"""
Board.__init__(self)
self.width = width
self.height = height
self.front = [width/2 for _ in range(height)]
self.positions = {}
# place front in positions
for y, x in enumerate(self.front):
self.positions[(x, y)] = 'f'
# place players
for n in range(players_num):
player = Soldier(0, n, 'left')
self.add_player(player, 'left')
player = Soldier(width-1, height-1-n, 'right')
self.add_player(player, 'right')
def test_a_cell_without_neighbours_dies_after_tick(self):
dying_cell = Cell(4, 4)
my_board = Board([self.fst_cell, self.snd_cell, self.trd_cell, dying_cell])
my_board.tick()
self.assertFalse(my_board.cell_alive_at(4, 4))
self.assertTrue(my_board.cell_alive_at(6, 6))
self.assertTrue(my_board.cell_alive_at(5, 6))
self.assertTrue(my_board.cell_alive_at(5, 7))
def test_get_range(self):
start = [
1, 2, 3, 4,
5, 6, 7, 8,
9, 10, 11, 12,
13, 14, 15, 16
]
row_first = [1, 2, 3, 4]
row_last = [13, 14, 15, 16]
col_first = [1, 5, 9, 13]
col_last = [4, 8, 12, 16]
board = Board(4, 4, start)
self.assertEqual(row_first, board._Board__get_range(0, 'left'))
self.assertEqual(row_first, board._Board__get_range(0, 'right'))
self.assertEqual(row_last, board._Board__get_range(3, 'left'))
self.assertEqual(row_last, board._Board__get_range(3, 'right'))
self.assertEqual(col_first, board._Board__get_range(0, 'up'))
self.assertEqual(col_first, board._Board__get_range(0, 'down'))
self.assertEqual(col_last, board._Board__get_range(3, 'up'))
self.assertEqual(col_last, board._Board__get_range(3, 'down'))
def py_get_liberty(matrix):
black_liberty = np.zeros((19, 19, 8), dtype=np.uint8)
white_liberty = np.zeros((19, 19, 8), dtype=np.uint8)
visited = {}
for i in range(19):
for j in range(19):
if matrix[i][j] == 1 and (i, j) not in visited:
groups = Board.get_group(i, j, matrix, visited=visited)
num_liberty = Board.check_liberty(groups, matrix, cnt=True)
if num_liberty > 8:
num_liberty = 8
for stone in groups:
black_liberty[stone[0]][stone[1]][num_liberty-1] = 1
if matrix[i][j] == 2 and (i, j) not in visited:
groups = Board.get_group(i, j, matrix, visited=visited)
num_liberty = Board.check_liberty(groups, matrix, cnt=True)
if num_liberty > 8:
num_liberty = 8
for stone in groups:
white_liberty[stone[0]][stone[1]][num_liberty-1] = 1
stones = np.concatenate((black_liberty, white_liberty), axis=2)
return stones
def test_a_died_cell_with_3_alive_neighbours_resurrects(self):
my_board = Board([self.fst_cell, self.snd_cell, self.trd_cell])
my_board.tick()
self.assertTrue(my_board.cell_alive_at(6, 6))
self.assertTrue(my_board.cell_alive_at(5, 6))
self.assertTrue(my_board.cell_alive_at(5, 7))
# Having 3 alive neighbours comes back to life.
self.assertTrue(my_board.cell_alive_at(6, 7))
def test_a_cell_with_more_then_3_neighbours_dies(self):
my_board = Board([self.fst_cell, self.snd_cell,
self.trd_cell, self.fourth_cell, self.fifth_cell])
my_board.tick()
# dying cells: first and fifth.
self.assertFalse(my_board.cell_alive_at(6, 6))
self.assertFalse(my_board.cell_alive_at(7, 5))
self.assertTrue(my_board.cell_alive_at(5, 6))
self.assertTrue(my_board.cell_alive_at(5, 7))
self.assertTrue(my_board.cell_alive_at(6, 7))
def check(self):
in_set = self.board.has_set()
# 1. Add up to 12 cards
if len(self.board.table) < TABLE_SIZE:
self.board = self.board.add_cards()
# 2. Continue adding cards until there's a set
elif not in_set and len(self.board.table) < TABLE_LIMIT:
self.board = self.board.add_cards()
# 3. Reshuffle if no set and window can't handle more cards
elif not in_set and len(self.board.table) == TABLE_LIMIT:
cards = random.shuffle(self.board.deck + self.board.table)
new_deck = cards[TABLE_SIZE:]
new_table = cards[:TABLE_SIZE]
self.board = Board(new_deck, new_table)
# 4. Break if added more than window can handle
assert(len(self.board.table) <= TABLE_LIMIT)
def client():
s = socket.socket()
host = HOST
print('connecting to ' + host)
port = PORT
s.connect((host, port))
game_play = GamePlay(
policy_net_path='../trained_models/policy',
value_net_path='../trained_models/value')
game = Game(handicap=HANDICAP)
while True:
message = s.recv(4096)
message = message.decode('utf-8')
print(type(message), message)
# board_mtx, next_to_move = message.split('|')
move, next_to_move, current_board, is_search = message.split('|')
print(move, next_to_move, current_board, is_search)
if int(is_search) == 1:
game_play.mcts.time_limit = 20
else:
game_play.mcts.time_limit = 0.5
while Board.mtx2str(game.boards[-1].board_mtx) != current_board:
print('roll_back')
game.roll_back()
print(len(game.boards))
moves = move.split(',')
x, y = int(moves[0][1:]), int(moves[1][1:-1])
game.mk_move(x, y)
output = game_play.play(game)
game.mk_move(output[0], output[1])
s.send(bytes(str(output), encoding='utf-8'))
# s.send(str(output))
s.close()
from game import Board
import random
import player
import pandas as pd
inp1 = -1
inp1 = -1
#For simulating data
for _ in range(500):
out = []
print("Game")
print(_)
win_percent_history = []
board = Board(False, False)
board.build_board()
win_percent_history = player.simulate_game(board, -1, 10)
for sim in win_percent_history:
for turn in sim:
for play in turn:
out.append(play)
df = pd.DataFrame(out, columns=['player','board', 'win_percent'])
df.to_csv('game_data2.csv', mode='a' ,index=False)
def __init__(self):
self.board = Board()
self.draw()
class Ui:
"""Command line user interface
Manages all the interaction with the user.
"""
color_dict = {0: "White", 1: "Black"}
piece_type_dict = {
0: "\u2654\u265a",
1: "\u2655\u265b",
2: "\u2656\u265c",
3: "\u2657\u265d",
4: "\u2658\u265e",
5: "\u2659\u265f"
}
x_str2int = {
"a": 0,
"b": 1,
"c": 2,
"d": 3,
"e": 4,
"f": 5,
"g": 6,
"h": 7
}
y_str2int = {
"1": 0,
"2": 1,
"3": 2,
"4": 3,
"5": 4,
"6": 5,
"7": 6,
"8": 7
}
x_int2str = {
0: "a",
1: "b",
2: "c",
3: "d",
4: "e",
5: "f",
6: "g",
7: "h"
}
y_int2str = {
0: "1",
1: "2",
2: "3",
3: "4",
4: "5",
5: "6",
6: "7",
7: "8"
}
def __init__(self):
self.board = Board()
self.draw()
def draw(self):
"""Draws the board configuration in the terminal"""
display = " a b c d e f g h \n" + \
"________________________________ \n "
# Loop over all x and y indices
for j in range(8):
display += " " + str(j + 1) + "|"
for i in range(8):
# Find the piece index for position [i, j]
position_ij = Position(i, j)
piece = self.board.find_piece(position_ij)
if piece.color != None:
display += " " + \
Ui.piece_type_dict[piece.type][piece.color] + " "
else:
# Draw an empty cell
display += " - "
# New line for different i value
display += "|" + str(j + 1) + " \n "
display += "_______________________________ \n" + \
" a b c d e f g h \n"
self.board_string = display
print(display)
def turn(self):
""""Performs a turn within ui"""
selected_piece = self.select_piece()
position = self.select_move(selected_piece)
self.board.turn(selected_piece, position, self.draw, self.check,
self.check_mate)
def select_piece(self) -> Piece:
"""Asks the user to select a piece to make a move with"""
question = Ui.color_dict[self.board.turn_color] + \
", your turn! Please select a piece. \n"
piece = Piece(None, None, None, None, None)
while piece.color is None or piece.color != self.board.turn_color:
coordinate = input(question)
position = self.coordinate2position(coordinate)
piece = self.board.find_piece(position)
question = "No piece of yours at this field, try again! \n"
return piece
def select_move(self, selected_piece: int) -> Position:
"""Asks the user where to move the selected piece"""
question = "The selected piece can move to " + \
self.moves2text(selected_piece) + "\n"
coordinate = input(question)
position = self.coordinate2position(coordinate)
while not position in selected_piece.moves:
question = "Your piece can't move to the selected field, try again! \n"
coordinate = input(question)
position = self.coordinate2position(coordinate)
return position
def moves2text(self, selected_piece: Piece) -> str:
"""Turns a list of positions into a string with coordinates"""
text = ""
for move in selected_piece.moves:
text += self.position2coordinate(move) + ", "
return text
def coordinate2position(self, coordinate: str) -> Position:
"""Converts user input to a board position"""
x = Ui.x_str2int[coordinate[0]]
y = Ui.y_str2int[coordinate[1]]
return Position(x, y)
def position2coordinate(self, position: Position) -> str:
"""Converts user a position to a ui coordinate"""
return Ui.x_int2str[position.x] + Ui.y_int2str[position.y]
def check(self):
"""Function that notifies players when check"""
print('Check!')
def check_mate(self):
"""Function that notifies players when check mate"""
print('Check mate! The game is over')
def test_print_action_boom(self):
string_action = Board().string_action
self.assertEqual(string_action([BOOM, (0, 0)]), "BOOM at (0, 0).")
self.assertEqual(string_action([BOOM, (5, 2)]), "BOOM at (5, 2).")
def test_move_row_works(self):
b = Board(seed=False)
# Make the top row 2 1 0 1
b._tiles[0, :] = [2, 1, 0, 1]
# Move to the left, so look at row left-to-right
row_1 = ((0, 0), (0, 1), (0, 2), (0, 3))
b._move_tile_row(row_1)
self.assertEqual(b._tiles.take(0), 2)
self.assertEqual(b._tiles.take(1), 2)
b._move_tile_row(row_1)
self.assertEqual(b._tiles.take(0), 4)
# Make the second row 1 1 0 2
b._tiles[1, :] = [1, 1, 0, 2]
row_2 = ((1, 0), (1, 1), (1, 2), (1, 3))
b._move_tile_row(row_2)
# Expect 2 2 0 0
self.assertEqual(b._tiles.take(4), 2)
self.assertEqual(b._tiles.take(5), 2)
b._move_tile_row(row_2)
self.assertEqual(b._tiles.take(4), 4)
# First column is now 2 2 0 0
# Move up to merge
col_1 = ((0, 0), (1, 0), (2, 0), (3, 0))
b._move_tile_row(col_1)
self.assertEqual(b._tiles.take(0), 8)
cnt = 0
for board_mtx, move in zip(games[0], games[1]):
cnt += 1
if cnt % 200 != 0:
continue
mtx = board_mtx
tic = time.time()
for i in range(20):
py_ret_mtx = py_get_liberty(mtx)
toc = time.time()
print(toc - tic)
tic = time.time()
for i in range(20):
string = Board.mtx2str(mtx)
string = gofeat.get_liberty(string)
ret_mtx = np.fromstring(string, sep=' ',
dtype=np.int).reshape(16, 19,
19).transpose(1, 2, 0)
toc = time.time()
print(toc - tic)
print(np.sum(py_ret_mtx - ret_mtx))
print(ret_mtx.shape)
# for i in range(16):
# print('num', i+1)
# li_b = Board(board_mtx=ret_mtx[i, :, :])
# li_canvas = li_b.visualize_board(grid_size=35)
# board = Board(board_mtx=board_mtx)
# canvas = board.visualize_board(grid_size=35)
def run(width, agent1, agent2, file1, file2, start, rounds):
n = 5
height = width
try:
board = Board(width=width, height=height, n_in_row=n)
game = Game(board)
# ############### human VS AI ###################
def parse_agent(agent_type, filename):
if agent_type == 'mcts_a0':
model_file = 'best_policy_8_8_5.model'
if filename:
model_file = filename
# load the trained policy_value_net in either Theano/Lasagne, PyTorch or TensorFlow
# best_policy = PolicyValueNet(width, height, model_file = model_file)
# mcts_player = MCTSPlayer(best_policy.policy_value_fn, c_puct=5, n_playout=400)
# load the provided model (trained in Theano/Lasagne) into a MCTS player written in pure numpy
try:
policy_param = pickle.load(open(model_file, 'rb'))
except:
policy_param = pickle.load(
open(model_file,
'rb'), encoding='bytes') # To support python3
best_policy = PolicyValueNetNumpy(width, height, policy_param)
player = MCTSPlayer(
best_policy.policy_value_fn, c_puct=5, n_playout=400
) # set larger n_playout for better performance
elif agent_type == 'mcts_pure':
player = MCTS_Pure(c_puct=5, n_playout=1000)
elif agent_type == 'minmax':
player = Minimax()
elif agent_type == 'dqn':
model_file = 'output/v_1/epoch_100/agent_2.pkl'
if filename:
model_file = filename
player = DQNPlayer(model_file)
elif agent_type == 'human':
player = Human()
else:
player = Human()
print('Illegal Agent Type. Defaulting to human player.')
return player
player1 = parse_agent(agent1, file1)
player2 = parse_agent(agent2, file2)
winners = []
for i in range(rounds):
winner = game.start_play(player1,
player2,
start_player=start - 1,
is_shown=1)
winners.append(winner)
winrate1 = winners.count(1) / rounds
winrate2 = winners.count(2) / rounds
print('Winners: ' + ','.join([str(w) for w in winners]))
print(str(agent1) + ' 1' + ' win rate: ' + str(winrate1))
print(str(agent2) + ' 2' + ' win rate: ' + str(winrate2))
except KeyboardInterrupt:
print('\n\rquit')
class Gui():
"""Grafical user interface for playing chess"""
font = 'Courier 20'
color_dict = {0: '#b0b0b0', 1: '#ffffff'}
piece_type_dict = {
0: {
0: '\u2654',
1: '\u265a'
},
1: {
0: '\u2655',
1: '\u265b'
},
2: {
0: '\u2656',
1: '\u265c'
},
3: {
0: '\u2657',
1: '\u265d'
},
4: {
0: '\u2658',
1: '\u265e'
},
5: {
0: '\u2659',
1: '\u265f'
}
}
turn_color_dict = {0: 'White', 1: 'Black'}
def __init__(self):
# Init board
self.board = Board()
# Init root
self.root = tk.Tk()
# Create general structure
self.board_frame = tk.Frame(self.root)
self.board_frame.pack()
self.test_frame = tk.Label(self.root, text='Welcome', font=Gui.font)
self.test_frame.pack()
self.user_input = tk.Entry(self.root, font=Gui.font)
self.user_input.configure(state='readonly')
self.user_input.pack()
self.user_input_given = tk.IntVar(master=self.user_input,
name='piece_type',
value=-1)
# Create buttons/fields
self.buttons = [[], [], [], [], [], [], [], []]
self.fields = [[], [], [], [], [], [], [], []]
for x, y in product(range(8), range(8)):
field_color = (x + y) % 2
self.fields[x].append(
tk.Frame(self.board_frame,
height=50,
width=50,
background=Gui.color_dict[field_color]))
self.fields[x][y].propagate(False)
self.fields[x][y].grid(column=x, row=8 - y)
self.buttons[x].append(
tk.Button(self.fields[x][y],
background=Gui.color_dict[field_color],
activebackground='#f2ff00',
borderwidth=0,
font='Courier 30'))
self.buttons[x][y].pack(fill='both', expand=True)
self.draw()
self.select_piece()
self.root.mainloop()
def ask_promotion_type(self):
"""Asks the user which piece to promote"""
self.user_input.bind('<Return>', self.promote2input)
self.test_frame.configure(text='Promote to type:')
self.user_input.configure(state='normal')
self.reset_buttons()
self.user_input_given.set(-1)
self.user_input.wait_variable(name='piece_type')
user_input = self.user_input_given.get()
self.user_input.delete(0, len(self.user_input.get()))
self.user_input.configure(state='readonly')
return user_input
def promote2input(self, event) -> int:
"""Gets the entered text from the entry box"""
type_dict = {
'king': 0,
'queen': 1,
'rook': 2,
'bishop': 3,
'knight': 4,
'pawn': 5
}
promotion_type = type_dict[self.user_input.get()]
# self.board.promote(piece, promotion_type)
self.user_input.bind('<Return>')
self.user_input_given.set(promotion_type)
def select_piece(self):
"""Select piece to move"""
color = self.board.turn_color
for x, rows in enumerate(self.buttons):
for y, button in enumerate(rows):
piece = self.board.find_piece(Position(x, y))
if piece.color == color and \
piece.moves != [] and \
piece.moves != None:
func = partial(self.show_moves, piece)
button.configure(command=func)
def show_moves(self, piece: Piece):
"""Marks the fields where the selected piece can move to"""
self.reset_buttons()
for move in piece.moves:
self.buttons[move.x][move.y].configure(background='#f2ff00',
command=partial(
self.select_move, piece,
move))
def select_move(self, piece: Piece, position):
"""Runs when player selects where to move to"""
self.reset_buttons()
self.board.recalculate(piece, position, self.ask_promotion_type)
self.board.delete_self_check()
self.board.turn_counter += 1
self.board.turn_color = int(not self.board.turn_color)
self.draw()
if self.board.check == True:
self.board.check_mate = \
self.board.ischeckmate(self.board.turn_color)
if self.board.check_mate == True:
self.test_frame.configure(text='Check mate!')
else:
self.test_frame.configure(text='Check!')
else:
message = Gui.turn_color_dict[self.board.turn_color] + \
', it\'s your turn'
self.test_frame.configure(text=message)
self.select_piece()
def draw(self):
"""Draws pieces on the board"""
for (x, y) in product(range(8), range(8)):
piece = self.board.find_piece(Position(x, y))
if piece.color != None:
self.buttons[x][y].config(
text=Gui.piece_type_dict[piece.type][piece.color])
else:
self.buttons[x][y].config(text='')
def reset_buttons(self):
"""Resets the buttons colors and commands"""
for x, y in product(range(8), range(8)):
button = self.buttons[x][y]
button.configure(command=False,
background=Gui.color_dict[(x + y) % 2])
import os
from game import Board
def bad_coordinates():
os.system("say 'bad coordinates'")
print("bad coordinates")
board = Board(4, 5, 6)
while (not board.game_over):
try:
os.system("say 'enter x coordinate'")
x = int(input("enter x coordinate: "))
os.system("say 'enter y coordinate'")
y = int(input("enter y coordinate: "))
except:
bad_coordinates()
if (x >= 0 and x < board.width and y >= 0 and y < board.height):
board.guess_square(x, y)
else:
bad_coordinates()
class TestCases(unittest.TestCase):
def setUp(self):
self.test_board1 = Board()
# The 11 white stack can't move, extra 1 white at (0, 0)
self.test_board2 = Board({
'white': [[11, 4, 4], [1, 0, 0]],
'black': [[1, 0, 4], [1, 1, 4], [1, 2, 4], [1, 3, 4], [1, 5, 4],
[1, 6, 4], [1, 7, 4], [1, 4, 0], [1, 4, 1], [1, 4, 2],
[1, 4, 3], [1, 4, 5], [1, 4, 6], [1, 4, 7]]
})
# White stack in corner that can move
self.test_board3 = Board({
'white': [[4, 7, 7]],
'black': [[3, 7, 6], [3, 6, 7], [2, 7, 5], [2, 5, 7], [1, 7, 4],
[1, 4, 7]]
})
# Board full of white pieces
self.test_board4 = Board({
'white': [[1, x, y] for x in range(8) for y in range(8)],
'black': []
})
def test_board_initilisation(self):
self.assertEqual(self.test_board1.board, [[0] * 8] * 8)
self.assertEqual(
self.test_board2.board,
[[1, 0, 0, 0, -1, 0, 0, 0], [0, 0, 0, 0, -1, 0, 0, 0],
[0, 0, 0, 0, -1, 0, 0, 0], [0, 0, 0, 0, -1, 0, 0, 0],
[-1, -1, -1, -1, 11, -1, -1, -1], [0, 0, 0, 0, -1, 0, 0, 0],
[0, 0, 0, 0, -1, 0, 0, 0], [0, 0, 0, 0, -1, 0, 0, 0]])
self.assertEqual(self.test_board2.n_white, 12)
self.assertEqual(self.test_board2.n_black, 14)
def test_get_all_white_actions(self):
self.assertEqual(
self.test_board1.get_all_white_actions(self.test_board1.board), [])
self.assertEqual(
self.test_board2.get_all_white_actions(self.test_board2.board),
[[BOOM,
(0, 0)], [1, (1, 0, 0, 1, 0)], [1,
(1, 0, 0, 0, 1)], [BOOM,
(4, 4)]])
self.assertEqual(
self.test_board3.get_all_white_actions(self.test_board3.board),
[[0, (7, 7)], [1, (1, 7, 7, 3, 7)], [1, (1, 7, 7, 7, 3)],
[1, (2, 7, 7, 3, 7)], [1, (2, 7, 7, 7, 3)], [1, (3, 7, 7, 3, 7)],
[1, (3, 7, 7, 7, 3)], [1, (4, 7, 7, 3, 7)], [1, (4, 7, 7, 7, 3)]])
def test_get_actions(self):
self.assertEqual(
self.test_board3.get_actions(self.test_board3.board, 7, 5),
[[0, (7, 5)], [1, (1, 7, 5, 6, 5)], [1, (1, 7, 5, 5, 5)],
[1, (1, 7, 5, 7, 6)], [1, (1, 7, 5, 7, 4)], [1, (1, 7, 5, 7, 3)],
[1, (2, 7, 5, 6, 5)], [1, (2, 7, 5, 5, 5)], [1, (2, 7, 5, 7, 6)],
[1, (2, 7, 5, 7, 4)], [1, (2, 7, 5, 7, 3)]])
def test_explode(self):
self.assertEqual(
self.test_board4.explode(self.test_board4.board, 0, 0),
([[0] * 8] * 8, 64))
self.assertEqual(
self.test_board2.explode(self.test_board2.board, 7, 4),
([[1] + [0] * 7] + [[0] * 8] * 7, 25))
def test_move(self):
self.assertEqual(
self.test_board3.move(self.test_board3.board, 1, 6, 7, 6, 5),
[[0] * 8] * 4 + [[0] * 7 + [-1], [0] * 7 + [-2],
[0] * 5 + [-1, 0, -2], [0] * 4 + [-1, -2, -3, 4]])
self.assertEqual(
self.test_board3.move(self.test_board3.board, 3, 6, 7, 4, 7),
[[0] * 8] * 4 + [[0] * 7 + [-4], [0] * 7 + [-2], [0] * 8,
[0] * 4 + [-1, -2, -3, 4]])
def test_print_action_boom(self):
string_action = Board().string_action
self.assertEqual(string_action([BOOM, (0, 0)]), "BOOM at (0, 0).")
self.assertEqual(string_action([BOOM, (5, 2)]), "BOOM at (5, 2).")
def test_print_action_move(self):
string_action = Board().string_action
self.assertEqual(string_action([MOVE, (1, 0, 0, 0, 1)]),
"MOVE 1 from (0, 0) to (0, 1).")
self.assertEqual(string_action([MOVE, (5, 2, 3, 7, 3)]),
"MOVE 5 from (2, 3) to (7, 3).")
def cycle():
startboard = np.array([[4] * 6, [0] * 2, [4] * 6]) # starting board
b = Board(startboard)
c = 1
player = 0
empty = [0] * 6
moves = 0
while b.board[0] != empty and b.board[2] != empty: # while at least one side of the board has stones in it
pos = r.randint(1, 6)
if c % 2 == 1:
player = 1
elif c % 2 == 0:
player = 2
if player == 1:
b.movep1(pos - 1)
moves+=1
while b.goagainp1:
b.movep1(pos - 1)
moves+=1
elif player == 2:
b.movep2(pos - 1)
moves+=1
while b.goagainp2:
b.movep2(pos - 1)
moves+=1
c += 1
if b.board[0] == empty:
b.board[1][1] += sum(b.board[2])
b.board[2] = empty
elif b.board[2] == empty:
b.board[1][0] += sum(b.board[0])
b.board[0] = empty
score = '{}-{}'.format(max(b.board[1]), min(b.board[1]))
if b.board[1].index(max(b.board[1])) == 0 and not (max(b.board[1]) == min(b.board[1])):
winner = 2
elif b.board[1].index(max(b.board[1])) == 1 and not (max(b.board[1]) == min(b.board[1])):
winner = 1
else:
winner = 0
'''
print(b)
if winner == 0:
print("It's a draw! Final score " + score)
else:
print("Player {} won! Final score ".format(winner) + score)
'''
return winner, moves
def __init__(self, init_model=None):
# params of the board and the game
self.board_width = 15
self.board_height = 15
self.n_in_row = 5
self.board = Board(width=self.board_width,
height=self.board_height,
n_in_row=self.n_in_row)
self.game = Game(self.board)
self.manual = Manual(self.board)
# training params
self.learn_rate = 1e-3
self.lr_multiplier = 1.0 # adaptively adjust the learning rate based on KL
self.temp = 1.0 # the temperature param
self.n_playout = 100 # num of simulations for each move
self.c_puct = 1
self.buffer_size = 100000
self.batch_size = 512 # mini-batch size for training
self.data_buffer = deque(maxlen=self.buffer_size)
self.play_batch_size = 1
self.epochs = 5 # num of train_steps for each update
self.episode_len = 0
self.kl_targ = 0.02
self.check_freq = 1
self.game_batch_num = 5
self.best_win_ratio = 0.55
# num of simulations used for the pure mcts, which is used as
# the opponent to evaluate the trained policy
self.pure_mcts_playout_num = 1000
self.lock = threading.Lock()
if init_model:
# start training from an initial policy-value net
self.g1 = tf.Graph()
with self.g1.as_default():
self.policy_value_net = PolicyValueNet(self.board_width,
self.board_height,
model_file=init_model,
graph=self.g1,
output='/data/data/')
# tf.reset_default_graph()
self.g2 = tf.Graph()
with self.g2.as_default():
self.policy_value_net_train = PolicyValueNet(self.board_width,
self.board_height,
model_file=init_model,
graph=self.g2,
output='/data/output/')
else:
# start training from a new policy-value net
self.g1 = tf.Graph()
with self.g1.as_default():
self.policy_value_net = PolicyValueNet(self.board_width,
self.board_height,
graph=self.g1,
output='./data/')
# tf.reset_default_graph()
self.g2 = tf.Graph()
with self.g2.as_default():
self.policy_value_net_train = PolicyValueNet(self.board_width,
self.board_height,
graph=self.g2,
output='./output/')
self.mcts_player = MCTSPlayer(self.policy_value_net.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.n_playout,
is_selfplay=1)
class TrainPipeline():
def __init__(self, init_model=None):
# params of the board and the game
self.board_width = 15
self.board_height = 15
self.n_in_row = 5
self.board = Board(width=self.board_width,
height=self.board_height,
n_in_row=self.n_in_row)
self.game = Game(self.board)
self.manual = Manual(self.board)
# training params
self.learn_rate = 1e-3
self.lr_multiplier = 1.0 # adaptively adjust the learning rate based on KL
self.temp = 1.0 # the temperature param
self.n_playout = 100 # num of simulations for each move
self.c_puct = 1
self.buffer_size = 100000
self.batch_size = 512 # mini-batch size for training
self.data_buffer = deque(maxlen=self.buffer_size)
self.play_batch_size = 1
self.epochs = 5 # num of train_steps for each update
self.episode_len = 0
self.kl_targ = 0.02
self.check_freq = 1
self.game_batch_num = 5
self.best_win_ratio = 0.55
# num of simulations used for the pure mcts, which is used as
# the opponent to evaluate the trained policy
self.pure_mcts_playout_num = 1000
self.lock = threading.Lock()
if init_model:
# start training from an initial policy-value net
self.g1 = tf.Graph()
with self.g1.as_default():
self.policy_value_net = PolicyValueNet(self.board_width,
self.board_height,
model_file=init_model,
graph=self.g1,
output='/data/data/')
# tf.reset_default_graph()
self.g2 = tf.Graph()
with self.g2.as_default():
self.policy_value_net_train = PolicyValueNet(self.board_width,
self.board_height,
model_file=init_model,
graph=self.g2,
output='/data/output/')
else:
# start training from a new policy-value net
self.g1 = tf.Graph()
with self.g1.as_default():
self.policy_value_net = PolicyValueNet(self.board_width,
self.board_height,
graph=self.g1,
output='./data/')
# tf.reset_default_graph()
self.g2 = tf.Graph()
with self.g2.as_default():
self.policy_value_net_train = PolicyValueNet(self.board_width,
self.board_height,
graph=self.g2,
output='./output/')
self.mcts_player = MCTSPlayer(self.policy_value_net.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.n_playout,
is_selfplay=1)
def get_equi_data(self, play_data):
"""augment the data set by rotation and flipping
play_data: [(state, mcts_prob, winner_z), ..., ...]
"""
extend_data = []
for state, mcts_porb, winner in play_data:
for i in [1, 2, 3, 4]:
# rotate counterclockwise
equi_state = np.array([np.rot90(s, i) for s in state])
equi_mcts_prob = np.rot90(np.flipud(
mcts_porb.reshape(self.board_height, self.board_width)), i)
extend_data.append((equi_state,
np.flipud(equi_mcts_prob).flatten(),
winner))
# flip horizontally
equi_state = np.array([np.fliplr(s) for s in equi_state])
equi_mcts_prob = np.fliplr(equi_mcts_prob)
extend_data.append((equi_state,
np.flipud(equi_mcts_prob).flatten(),
winner))
return extend_data
def collect_selfplay_data(self, n_games=1):
"""collect self-play data for training"""
for i in range(n_games):
# self.lock.acquire()
# print("game {}".format(i))
with self.g1.as_default():
'''mcts_player = MCTSPlayer(self.policy_value_net.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.n_playout,
is_selfplay=1)
board = Board(width=self.board_width,
height=self.board_height,
n_in_row=self.n_in_row)
game = Game(board)'''
winner, play_data = self.game.start_self_play(self.mcts_player,
is_shown=0,
temp=self.temp)
# self.lock.release()
play_data = list(play_data)[:]
self.episode_len = len(play_data)
# augment the data
play_data = self.get_equi_data(play_data)
self.data_buffer.extend(play_data)
# print("self play end...")
def collect_manual_data(self, file):
winner, play_data = self.manual.read_manual_data(file)
# read the chess manual fail
if winner == 0:
return
play_data = list(play_data)[:]
self.episode_len = len(play_data)
# augment the data
play_data = self.get_equi_data(play_data)
self.data_buffer.extend(play_data)
def collect_test_data(self):
self.board.init_board()
states, mcts_probs, current_players = [], [], []
move = 128
self.board.do_move(112)
states.append(self.board.current_state())
probs = np.zeros(self.board.width * self.board.height)
probs[[move]] = 1
mcts_probs.append(probs)
current_players.append(self.board.current_player)
winners_z = np.array([1])
play_data = zip(states, mcts_probs, winners_z)
play_data = list(play_data)[:]
self.data_buffer.extend(play_data)
def policy_update(self):
"""update the policy-value net"""
mini_batch = random.sample(self.data_buffer, self.batch_size)
state_batch = [data[0] for data in mini_batch]
mcts_probs_batch = [data[1] for data in mini_batch]
winner_batch = [data[2] for data in mini_batch]
with self.g2.as_default():
for i in range(self.epochs):
loss, entropy = self.policy_value_net_train.train_step(
state_batch,
mcts_probs_batch,
winner_batch,
self.learn_rate*self.lr_multiplier)
print((
"lr_multiplier:{:.3f},"
"loss:{},"
"entropy:{},"
).format(
self.lr_multiplier,
loss,
entropy))
return loss, entropy
def policy_evaluate(self, n_games=10):
"""
Evaluate the trained policy by playing against the pure MCTS player
Note: this is only for monitoring the progress of training
"""
print("evaluating...")
current_mcts_player = MCTSPlayer(self.policy_value_net_train.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.pure_mcts_playout_num)
best_mcts_player = MCTSPlayer(self.policy_value_net.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.pure_mcts_playout_num)
win_cnt = defaultdict(int)
for i in range(n_games):
winner = self.game.start_play(current_mcts_player,
best_mcts_player,
start_player=i % 2,
is_shown=0)
win_cnt[winner] += 1
win_ratio = 1.0*(win_cnt[1] + 0.5*win_cnt[-1]) / n_games
print("num_playouts:{}, win: {}, lose: {}, tie:{}".format(
self.pure_mcts_playout_num,
win_cnt[1], win_cnt[2], win_cnt[-1]))
# save the current_model
self.policy_value_net_train.save_model('/data/output/current_policy.model')
if win_ratio > self.best_win_ratio:
print("New best policy!!!!!!!!")
# update the best_policy
self.policy_value_net_train.save_model('/data/output/best_policy.model')
self.g1 = tf.Graph()
with self.g1.as_default():
self.policy_value_net = PolicyValueNet(self.board_width,
self.board_height,
model_file='/data/output/best_policy.model',
graph=self.g1,
output='/data/data/')
return win_ratio
def run(self):
"""run the training pipeline"""
try:
'''coord = tf.train.Coordinator()
self_play = [threading.Thread(target=self.collect_selfplay_data, args=(self.play_batch_size,)) for i in range(4)]
for sp in self_play:
sp.start()
coord.join(self_play)
while len(self.data_buffer) < self.batch_size:
print(len(self.data_buffer))
time.sleep(3)
pass'''
multiplier = [0.1, 0.1, 0.01, 0.01, 0.01]
step = 0
for n in range(self.game_batch_num):
self.collect_selfplay_data(self.play_batch_size)
# self.collect_test_data()
self.policy_value_net.n_step += 1
print("batch i:{}, episode_len:{}".format(
self.policy_value_net.n_step, self.episode_len))
# optimisation
if len(self.data_buffer) > self.batch_size:
for i in range(100):
self.policy_update()
# evaluation
if self.policy_value_net.n_step % self.check_freq == 0:
# self.lr_multiplier = multiplier[step]
# step += 1
self.mcts_player.mcts._discount = 1 - 0.98*(1 - self.mcts_player.mcts._discount)
print("current self-play batch: {}, discount: {}".format(
self.policy_value_net.n_step, self.mcts_player.mcts._discount))
# self.lock.acquire()
self.policy_evaluate(n_games=15)
# self.lock.release()
except KeyboardInterrupt:
print('\n\rquit')
def game():
# For screen-casting
if RECORD_SCREEN:
recorder = ScreenRecorder(WIDTH, HEIGHT, FPS, out_file=OUTPUT_FILE)
run = True
board = Board.from_start_position()
solver = AutoSolver(board)
selected_piece = None
# A surface to draw the board onto..
board_surf = pygame.Surface(BOARD_SIZE)
loader = Loader()
def draw():
board_color = (205, 127, 50)
text_background = (0, 100, 255)
text_color = (255, 255, 255)
# Fill the window and the board
win.fill(darken_color(board_color, 0.5))
board_surf.fill(board_color)
# Draw the title label onto the window
pygame.draw.rect(win, text_background, (TITLE_OFFSETS, TITLE_SIZE))
title_label = main_font.render(f"KLOTSKI PUZZLE", 1, text_color)
win.blit(title_label,
(TITLE_OFFSETS[0] + TITLE_SIZE[0] // 2 -
title_label.get_width() // 2, TITLE_OFFSETS[1] +
TITLE_SIZE[1] // 2 - title_label.get_height() // 2))
# Draw the steps label onto the window
pygame.draw.rect(win, text_background, (SCORE_OFFSETS, SCORE_SIZE))
steps_label = main_font.render(f"Step {board.number_of_steps}", 1,
text_color)
win.blit(steps_label,
(SCORE_OFFSETS[0] + SCORE_SIZE[0] // 2 -
steps_label.get_width() // 2, SCORE_OFFSETS[1] +
SCORE_SIZE[1] // 2 - steps_label.get_height() // 2))
# Draw the board and copy it onto the window
board.draw(board_surf, TILE_SIZE)
win.blit(board_surf, BOARD_OFFSETS)
if board.is_solved:
# Show the message when game is solved
# NOTE: Game does not end when puzzle is solved, user can continue..
success_label = main_font.render(f"Congratulations!", 1,
text_color)
win.blit(success_label,
(BOARD_OFFSETS[0] + BOARD_SIZE[0] // 2 -
success_label.get_width() // 2, BOARD_OFFSETS[1] +
BOARD_SIZE[1] // 2 - success_label.get_height() // 2))
if solver.loading:
# Show a loader when auto-solver is computing the moves.
loader.draw(
win,
pygame.Rect(
(WIDTH // 2 - TILE_SIZE // 2, HEIGHT // 2 - TILE_SIZE // 2,
TILE_SIZE, TILE_SIZE)))
def handle_select(pos):
# Handles mouse button down event.
# Sets the selected_piece if a piece is selected
nonlocal selected_piece
selected_piece = None
pos = pos[0] - BOARD_OFFSETS[0], pos[1] - BOARD_OFFSETS[1]
if 0 <= pos[0] < BOARD_SIZE[0] and 0 <= pos[1] < BOARD_SIZE[1]:
position = Position(pos[0] // TILE_SIZE, pos[1] // TILE_SIZE)
selected_piece = board.get_piece(position)
def handle_drop(pos):
# Handles mouse button up event.
# Moves the selected_piece if to specified position if allowed.
# Specified position must be an empty position!
nonlocal selected_piece
pos = pos[0] - BOARD_OFFSETS[0], pos[1] - BOARD_OFFSETS[1]
if 0 <= pos[0] < BOARD_SIZE[0] and 0 <= pos[1] < BOARD_SIZE[1]:
click_position = Position(pos[0] // TILE_SIZE, pos[1] // TILE_SIZE)
if selected_piece:
possible_pos = board.can_move(selected_piece, click_position)
if possible_pos:
board.move(selected_piece, possible_pos)
def reset():
# creates a new board to reset it
nonlocal board, selected_piece, solver
board = Board.from_start_position()
selected_piece = None
# Reset the solver as well
solver = AutoSolver(board)
def handle_user_event(_event):
nonlocal selected_piece
if _event.type == pygame.KEYDOWN:
# Board reset
if _event.key == pygame.K_r:
reset()
# History events
if _event.key == pygame.K_LEFT:
board.history_back()
if _event.key == pygame.K_RIGHT:
board.history_forward()
# Solver
if _event.key == pygame.K_a: # Normal Solver
selected_piece = None
solver.enable()
if _event.key == pygame.K_s: # Fast solver
selected_piece = None
solver.enable(int(FPS * 0.1))
if _event.type == pygame.MOUSEBUTTONDOWN and _event.button == 1: # left click
handle_select(_event.pos)
if _event.type == pygame.MOUSEBUTTONUP and _event.button == 1: # left click
handle_drop(_event.pos)
while run:
draw()
pygame.display.update()
if RECORD_SCREEN:
recorder.capture_frame(win)
solver.loop()
for event in pygame.event.get():
if event.type == pygame.QUIT or \
(event.type == pygame.KEYDOWN and event.key == pygame.K_q):
run = False
if not solver.enabled:
# User inputs taken only when solver not running
handle_user_event(event)
if not solver.enabled:
# Power keys while navigating history
# Allows continuous press
keys = pygame.key.get_pressed()
if keys[pygame.K_DOWN]:
board.history_back()
elif keys[pygame.K_UP]:
board.history_forward()
clock.tick(FPS)
if RECORD_SCREEN:
recorder.stop()
pygame.quit()
def __init__(self):
"""
關於訓練的初始設置
*補充說明
kl 用於計算 lr (learning rate)
"""
# run() -----------------------------------------------------------------------------------
self.game_batch_num = -1 # 跑一次訓練的重複次數，負值代表不限制
self.play_batch_size = 1 # 自我訓練的執行次數
self.batch_size = 1024 # 每次要訓練的資料量，當 data_buffer 的資料累積到超過本數值就會更新 policy
self.check_freq = 50 # 每訓練 ( check_freq ) 次就會與MCTS比賽
self.save_freq = 50 # 每訓練 ( save_freq ) 次就會存檔
# collect_selfplay_data() -----------------------------------------------------------------
self.buffer_size = 10000
self.data_buffer = deque(maxlen=self.buffer_size)
self.kl_targ = 0.02
# policy_update() -------------------------------------------------------------------------
self.epochs = 5 # 每次更新的 epochs 數
# board -----------------------------------------------------------------------------------
self.board_width = 9 # 寬度
self.board_height = 9 # 高度
self.n_in_row = 5 # 多少顆連成一線獲得勝利
self.board = Board(width=self.board_width,
height=self.board_height,
n_in_row=self.n_in_row)
self.game = Game(self.board)
# keras -----------------------------------------------------------------------------------
self.learn_rate = 2e-3
self.lr_multiplier = 1.0 # 基於KL自適應調整學習率
self.temp = 1.0 # 溫度參數，太小會導致訓練不夠全面
file_folder = './n400-o'
model_tag = '9_9_5_o'
self.current_model= f'{file_folder}/current_model_{model_tag}.h5'
self.best_model= f'{file_folder}/best_model_{model_tag}.h5'
init_model = self.current_model
self.policy_value_net = PolicyValueNet(self.board_width,
self.board_height,
model_file = init_model if os.path.exists(init_model) else None)
self.progress = file_folder + '/progress.csv'
self.evaluate_path = file_folder + '/evaluate.csv'
self.history_path = file_folder + '/history.csv'
self.history = []
# MCTS ------------------------------------------------------------------------------------
self.c_puct = 5 # MCTS的搜索偏好
self.loss_goal = 0 #! 存檔時 loss 小於此值會增加訓練時的 n_playout 次數
self.pure_mcts_playout_num = 1000 # MCTS每一步的模擬次數，隨著模型強度提升
self.pure_mcts_playout_num_upgrade = 1000 # MCTS隨著模型強度提升的模擬次數
self.best_win_ratio = 0.0
self.n_playout = 400 # 神經網路每一步的模擬次數，越大代表結果越依賴MCTS的技巧，否則依靠神經網路的判斷
self.n_playout_training = 400
self.n_playout_growth = 0
self.n_playout_limit = 2000
self.MCTS_levelup()
def play():
board = Board()
A, B = Player(board, 'white'), Player(board, 'black')
class MctsTest():
def __init__(self, init_model=None):
# 棋盘大小 8*8, 5个子连起来
self.board_width = 8
self.board_height = 8
self.n_in_row = 5 # n子相连
self.policy_evaluate_size = 2 # 策略评估胜率时的模拟对局次数
self.batch_size = 1 # data_buffer中对战次数超过n次后开始启动模型训练
self.board = Board(width=self.board_width,
height=self.board_height,
n_in_row=self.n_in_row)
self.game = Game(self.board)
# training params
self.learn_rate = 2e-3
self.lr_multiplier = 1.0 # 基于KL的自适应学习率
self.temp = 1.0 # the temperature param
self.n_playout = 400 # 每个动作的模拟次数
self.c_puct = 5
self.buffer_size = 10000 # cache对战记录个数
self.data_buffer = deque(maxlen=self.buffer_size) # 完整对战历史记录，用于训练
self.epochs = 5 # 每次更新策略价值网络的训练步骤数
self.kl_targ = 0.02 # 策略价值网络KL值目标
self.best_win_ratio = 0.0
# 纯MCT的模拟数，用于评估策略模型
self.pure_mcts_playout_num = 5
self.policy_value_net = PolicyValueNet(self.board_width,
self.board_height)
# 创建使用策略价值网络来指导树搜索和评估叶节点的MCTS玩家
"""self.mcts_player = MCTSPlayer(self.policy_value_net.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.n_playout,
is_selfplay=1)"""
def get_equi_data(self, play_data):
"""
通过旋转和翻转增加数据集
play_data: [(state, mcts_prob, winner_z), ..., ...]
"""
extend_data = []
for state, mcts_porb, winner in play_data:
for i in [1, 2, 3, 4]:
# 逆时针旋转
equi_state = np.array([np.rot90(s, i) for s in state])
equi_mcts_prob = np.rot90(
np.flipud(
mcts_porb.reshape(self.board_height,
self.board_width)), i)
extend_data.append(
(equi_state, np.flipud(equi_mcts_prob).flatten(), winner))
# 水平翻转
equi_state = np.array([np.fliplr(s) for s in equi_state])
equi_mcts_prob = np.fliplr(equi_mcts_prob)
extend_data.append(
(equi_state, np.flipud(equi_mcts_prob).flatten(), winner))
return extend_data
def policy_update(self):
"""更新策略价值网络policy-value"""
# 随机抽取data_buffer中的对抗数据
mini_batch = random.sample(self.data_buffer, self.batch_size)
state_batch = [data[0] for data in mini_batch]
mcts_probs_batch = [data[1] for data in mini_batch]
winner_batch = [data[2] for data in mini_batch]
old_probs, old_v = self.policy_value_net.policy_value(state_batch)
# 训练策略价值网络
for i in range(self.epochs):
loss, entropy = self.policy_value_net.train_step(
state_batch, mcts_probs_batch, winner_batch,
self.learn_rate * self.lr_multiplier)
new_probs, new_v = self.policy_value_net.policy_value(state_batch)
kl = np.mean(
np.sum(old_probs *
(np.log(old_probs + 1e-10) - np.log(new_probs + 1e-10)),
axis=1))
if kl > self.kl_targ * 4: # 如果D_KL跑偏则尽早停止
break
# 自动调整学习率
if kl > self.kl_targ * 2 and self.lr_multiplier > 0.1:
self.lr_multiplier /= 1.5
elif kl < self.kl_targ / 2 and self.lr_multiplier < 10:
self.lr_multiplier *= 1.5
explained_var_old = (1 -
np.var(np.array(winner_batch) - old_v.flatten()) /
np.var(np.array(winner_batch)))
explained_var_new = (1 -
np.var(np.array(winner_batch) - new_v.flatten()) /
np.var(np.array(winner_batch)))
logging.info(
("TEST kl:{:.5f},"
"lr_multiplier:{:.3f},"
"loss:{},"
"entropy:{},"
"explained_var_old:{:.3f},"
"explained_var_new:{:.3f}").format(kl, self.lr_multiplier, loss,
entropy, explained_var_old,
explained_var_new))
return loss, entropy
def policy_evaluate(self, n_games=10):
"""
策略胜率评估：模型与纯MCTS玩家对战n局看胜率
"""
# AlphaGo Zero风格的MCTS玩家（使用策略价值网络来指导树搜索和评估叶节点）
current_mcts_player = MCTSPlayer(self.policy_value_net.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.n_playout)
# 纯MCTS玩家
pure_mcts_player = MCTSPurePlayer(c_puct=5,
n_playout=self.pure_mcts_playout_num)
win_cnt = defaultdict(int)
for i in range(n_games):
# 对战
winner = self.game.start_play(current_mcts_player,
pure_mcts_player,
start_player=i % 2,
is_shown=0)
win_cnt[winner] += 1
# 胜率
win_ratio = 1.0 * (win_cnt[1] + 0.5 * win_cnt[-1]) / n_games
logging.info("TEST Num_playouts:{}, win: {}, lose: {}, tie:{}".format(
self.pure_mcts_playout_num, win_cnt[1], win_cnt[2], win_cnt[-1]))
return win_ratio
def run(self):
"""启动训练"""
try:
#test
# 初始化棋盘
self.board.init_board()
print(self.board)
print(self.board.current_player)
print(self.board.availables)
print(self.board.states)
print(self.board.last_move)
p1, p2 = self.board.players
states, mcts_probs, current_players = [], [], []
# 纯MCTS玩家
#player = self.mcts_player
player = MCTSPurePlayer(c_puct=5,
n_playout=self.pure_mcts_playout_num)
print('------get_action------')
#move, move_probs = player.get_action(self.board, temp=self.temp, return_prob=1)
move = player.get_action(self.board)
print(move)
"""# 保存当前盘面
states.append(self.board.current_state())
current_players.append(self.board.current_player)
# 执行落子
print('------do_move------')
self.board.do_move(move)
self.game.graphic(self.board, p1, p2)
# 检查游戏是否结束
print('------check_game_end------')
end, winner = self.board.game_end()
if end:
# 从当前玩家视角确定winner
winners_z = np.zeros(len(current_players))
if winner != -1: # 不是和棋
winners_z[np.array(current_players) == winner] = 1.0 # 更新赢家步骤位置=1
winners_z[np.array(current_players) != winner] = -1.0 # 更新输家步骤位置=-1
# 重置MCTS根结点
player.reset_player()
if winner != -1:
print("Game end. Winner is player:", winner)
else:
print("Game end. Tie")
print(winner, zip(states, mcts_probs, winners_z))
"""
"""
i=0
# 1.收集自我对抗数据
# 使用MCTS蒙特卡罗树搜索进行自我对抗
winner, play_data = self.game.start_self_play(self.mcts_player, temp=self.temp)
play_data = list(play_data)[:]
self.episode_len = len(play_data)
print(play_data)
print(self.episode_len)
# 把翻转棋盘数据加到数据集里
play_data = self.get_equi_data(play_data)
# 保存对抗数据到data_buffer
self.data_buffer.extend(play_data)
logging.info("TEST Batch i:{}, episode_len:{}".format(i + 1, self.episode_len))
# 2.使用对抗数据重新训练策略价值网络模型
if len(self.data_buffer) >= self.batch_size:
loss, entropy = self.policy_update()
# 3.检查一下当前模型胜率
logging.info("TEST Current self-play batch: {}".format(i + 1))
# 策略胜率评估：模型与纯MCTS玩家对战n局看胜率
win_ratio = self.policy_evaluate(self.policy_evaluate_size)
self.policy_value_net.save_model(CUR_PATH + '/model/current_test_{}_{}.model'.format(self.board_width, self.board_height))
if win_ratio > self.best_win_ratio: # 胜率超过历史最优模型
logging.info("TEST New best policy!!!!!!!!batch:{} win_ratio:{}->{} pure_mcts_playout_num:{}".format(i + 1, self.best_win_ratio, win_ratio, self.pure_mcts_playout_num))
self.best_win_ratio = win_ratio
# 保存当前模型为最优模型best_policy
self.policy_value_net.save_model(CUR_PATH + '/model/best_test_{}_{}.model'.format(self.board_width, self.board_height))
# 如果胜率=100%，则增加纯MCT的模拟数
if (self.best_win_ratio == 1.0 and self.pure_mcts_playout_num < 5000):
self.pure_mcts_playout_num += 1000
self.best_win_ratio = 0.0
"""
except KeyboardInterrupt:
logging.info('\n\rquit')
game_input = input("game: ")
if(game_input == "n"):
game_num += 1
elif(game_input == "exit"):
break
else:
game_num = int(game_input)
turn_num = 0
game = all_data[game_num]
while(True):
turn_input = input("turn: ")
if(turn_input == "n" or turn_input == ""):
turn_num += 1
elif(turn_input == "b"):
break
else:
turn_num = int(turn_input)
states = game["states"]
turn = states[turn_num]
not_board = Board()
not_board.board[1:-1, 1:-1] = idtobits(np.array(turn["board"]))
not_board.print_board()
print(turn["player_turn"]-1)
print(turn["pieces_left"])
#interesting games:
#599
def test_new_board_starts_with_two_squares(self):
b = Board()
self.assertFalse(b.is_board_full())
self.assertEqual(len(b.empty_tiles()), len(b._tiles.flatten()) - 2)
def preprocess_board(board_mtx, y, random=True, contain_liberty=False):
# rand = np.random.randint(0, 2)
# if rand == 1:
# board_mtx = board_mtx.T
# y['next_to_move'] = (y['next_to_move'][1], y['next_to_move'][0])
# y['current_move'] = (y['current_move'][1], y['current_move'][0])
if random:
rand = np.random.randint(0, 8)
if rand <= 3:
board_mtx = board_mtx.T
y['current_move'] = (y['current_move'][1],
y['current_move'][0])
y['next_move'] = (y['next_move'][1], y['next_move'][0])
i = rand % 4
if i == 1:
board_mtx = np.rot90(board_mtx)
y['current_move'] = (18 - y['current_move'][1],
y['current_move'][0])
y['next_move'] = (18 - y['next_move'][1], y['next_move'][0])
# print(a[2-idx[1]][idx[0]])
if i == 2:
board_mtx = np.rot90(board_mtx)
board_mtx = np.rot90(board_mtx)
y['current_move'] = (18 - y['current_move'][1],
y['current_move'][0])
y['next_move'] = (18 - y['next_move'][1], y['next_move'][0])
y['current_move'] = (18 - y['current_move'][1],
y['current_move'][0])
y['next_move'] = (18 - y['next_move'][1], y['next_move'][0])
if i == 3:
board_mtx = np.rot90(board_mtx)
board_mtx = np.rot90(board_mtx)
board_mtx = np.rot90(board_mtx)
y['current_move'] = (18 - y['current_move'][1],
y['current_move'][0])
y['next_move'] = (18 - y['next_move'][1], y['next_move'][0])
y['current_move'] = (18 - y['current_move'][1],
y['current_move'][0])
y['next_move'] = (18 - y['next_move'][1], y['next_move'][0])
y['current_move'] = (18 - y['current_move'][1],
y['current_move'][0])
y['next_move'] = (18 - y['next_move'][1], y['next_move'][0])
black_stones = np.zeros((19, 19, 1), dtype=np.uint8)
black_stones[board_mtx == 1] = 1
white_stones = np.zeros((19, 19, 1), dtype=np.uint8)
white_stones[board_mtx == 2] = 1
if contain_liberty:
black_liberty = np.zeros((19, 19, 8), dtype=np.uint8)
white_liberty = np.zeros((19, 19, 8), dtype=np.uint8)
visited = {}
for i in range(19):
for j in range(19):
if board_mtx[i][j] == 1 and (i, j) not in visited:
groups = Board.get_group(i,
j,
board_mtx,
visited=visited)
num_liberty = Board.check_liberty(groups,
board_mtx,
cnt=True)
if num_liberty > 8:
num_liberty = 8
for stone in groups:
black_liberty[stone[0]][stone[1]][num_liberty -
1] = 1
if board_mtx[i][j] == 2 and (i, j) not in visited:
groups = Board.get_group(i,
j,
board_mtx,
visited=visited)
num_liberty = Board.check_liberty(groups,
board_mtx,
cnt=True)
if num_liberty > 8:
num_liberty = 8
for stone in groups:
white_liberty[stone[0]][stone[1]][num_liberty -
1] = 1
black_stones = np.concatenate((black_stones, black_liberty),
axis=2)
white_stones = np.concatenate((white_stones, white_liberty),
axis=2)
# for i in range(9):
# print(board_mtx)
# print('liberty:', i)
# print(black_stones[:, :, i])
# print('===')
#
# print('XXXXXX')
# print('XXXXXX')
# print('XXXXXX')
#
# for i in range(9):
# print(board_mtx)
# print('liberty:', i)
# print(white_stones[:, :, i])
# print('===')
# exit()
stones = np.concatenate((black_stones, white_stones), axis=2)
ones = np.ones((19, 19, 1), dtype=np.uint8)
last_move = np.zeros((19, 19, 1), dtype=np.uint8)
if not y['ko_state:']:
last_move[y['current_move'][0]][y['current_move'][1]] = 1
else:
last_move[y['current_move'][0]][y['current_move'][1]] = -1
is_black_next = np.ones((19, 19, 1), dtype=np.uint8)
if y['next_to_play'] == 2:
is_black_next -= 1
feat = np.concatenate((stones, last_move, is_black_next, ones), axis=2)
return feat
def test_can_make_moves(self):
b = Board(seed=False)
# Make sure it's fully empty first
self.assertEqual(len(b.empty_tiles()), 16)
# Put a square in the top left corner
val = 1
pos = 0
b._tiles.put(pos, val)
# Move in the negative direction along the rows (up)
# Then move in the negative direction along the columns (left)
# Should do nothing, we are at the edge
for d in (1, 3):
b.move(d, suppress_invalid=True)
self.assertEqual(b._tiles.take(pos), val)
# Move in the positive direction along the rows (down)
# This should move three rows
b.move(0)
pos += b.length * 3
self.assertEqual(b._tiles.take(pos), val)
# At the bottom - movement should do nothing
b.move(0, suppress_invalid=True)
self.assertEqual(b._tiles.take(pos), val)
# Move in the positive direction along the columns (right)
# This should move three columns
b.move(2)
pos += 3
self.assertEqual(b._tiles.take(pos), val)
# At the right - movement should do nothing
b.move(2, suppress_invalid=True)
self.assertEqual(b._tiles.take(pos), val)
def test_print_action_move(self):
string_action = Board().string_action
self.assertEqual(string_action([MOVE, (1, 0, 0, 0, 1)]),
"MOVE 1 from (0, 0) to (0, 1).")
self.assertEqual(string_action([MOVE, (5, 2, 3, 7, 3)]),
"MOVE 5 from (2, 3) to (7, 3).")
def test_board():
print("Board() class tests")
dimension = int(sys.argv[1])
board = Board(dimension)
board_2 = Board(dimension)
print("Imprimimos tablero vacio: ")
print(board)
board.update_board([0, 2], 'X')
board.update_board([0, 0], 'O')
board.update_board([1, 2], 'X')
board.update_board([2, 2], 'X')
board.update_board([1, 0], 'X')
board.update_board([2, 0], 'O')
board.update_board([0, 1], 'O')
board.update_board([1, 1], 'X')
board.update_board([2, 1], 'X')
if dimension == 4:
board.update_board([3, 3], 'X')
print("Imprimimos tablero con contenido: ")
print(board)
print(board.is_tateti())
print(board.get_board())
print(board.get_id())
print(board.get_dimension())
# board_2
print(board_2)
print(board_2.is_tateti())
board_2.update_board([0, 0], 'X')
print(board_2)
print(board_2.is_tateti())
def test_player():
print("Player() class tests")
dimension = 3
board_player = Board(dimension)
print("Imprimimos tablero vacio: ")
print(board_player)
board_player.update_board([0, 2], 'X')
board_player.update_board([0, 0], 'O')
board_player.update_board([1, 2], 'X')
board_player.update_board([2, 2], 'X')
board_player.update_board([1, 0], 'X')
board_player.update_board([2, 0], 'O')
board_player.update_board([0, 1], 'O')
board_player.update_board([1, 1], 'X')
#board_player.update_board([2, 1], 'X')
print(board_player)
player_1 = Player('Joaquin', 0, 0, 0)
player_2 = Player('Xano', 1, 1, 1)
print(player_1)
print(player_2)
player_1.movement(board_player)
print(board_player)
print(board_player.is_tateti())
def initUI(self):
self.tup = (None, None)
self.board = Board() # 棋盘类
self.board.init_board(1)
palette1 = QPalette() # 设置棋盘背景
palette1.setBrush(self.backgroundRole(),
QtGui.QBrush(QtGui.QPixmap('img/linesofaction.png')))
self.setPalette(palette1)
# self.setStyleSheet("board-image:url(img/chessboard.jpg)") # 不知道这为什么不行
self.setCursor(Qt.PointingHandCursor) # 鼠标变成手指形状
# self.sound_piece = QSound("sound/luozi.wav") # 加载落子音效
# self.sound_win = QSound("sound/win.wav") # 加载胜利音效
# self.sound_defeated = QSound("sound/defeated.wav") # 加载失败音效
self.resize(WIDTH, HEIGHT) # 固定大小 540*540
self.setMinimumSize(QtCore.QSize(WIDTH, HEIGHT))
self.setMaximumSize(QtCore.QSize(WIDTH, HEIGHT))
self.setWindowTitle("Lines-Of-Action") # 窗口名称
self.setWindowIcon(QIcon('img/black.png')) # 窗口图标
# self.lb1 = QLabel(' ', self)
# self.lb1.move(20, 10)
self.black = QPixmap('img/black.png')
self.white = QPixmap('img/white.png')
self.piece_now = BLACK # 黑棋先行
self.my_turn = True # 玩家先行
self.step = 0 # 步数
self.x, self.y = 1000, 1000
#self.mouse_point = LaBel(self) # 将鼠标图片改为棋子
# self.mouse_point.setScaledContents(True)
# self.mouse_point.setPixmap(self.black) # 加载黑棋
# self.mouse_point.setGeometry(270, 270, PIECE, PIECE)
self.pieces = [[
LaBel(self),
LaBel(self),
LaBel(self),
LaBel(self),
LaBel(self),
LaBel(self),
LaBel(self),
LaBel(self)
] for _ in range(8)] # 新建棋子标签，准备在棋盘上绘制棋子
# for piece in self.pieces:
# piece.setVisible(True) # 图片可视
# piece.setScaledContents(True) # 图片大小根据标签大小可变
for i in range(8):
for j in range(8):
self.pieces[i][j].setVisible(True)
self.pieces[i][j].setScaledContents(True)
#self.mouse_point.raise_() # 鼠标始终在最上层
self.ai_down = True # AI已下棋，主要是为了加锁，当值是False的时候说明AI正在思考，这时候玩家鼠标点击失效，要忽略掉 mousePressEvent
self.setMouseTracking(True)
self.DrawPieces()
self.show()
class App(ConnectionListener):
def __init__(self, host, port, run_server=False):
self._running = True
self._screen = None
self.reset_sound = None
self.run_server = run_server
self.size = self.width, self.height = 1800, 960
self.board = Board(self)
self.dice = Dice(self)
self.init_pieces()
self.player_count = 0
self.other_mouse = OtherMouse()
if self.run_server:
self.server = BackgammonServer(localaddr=(host, port))
self.Connect((host, port))
def init_pieces(self, send=True):
self.pieces = list()
self.fields = [[] for _ in range(24)]
self.fields[0] = [True] * 2
self.fields[5] = [False] * 5
self.fields[7] = [False] * 3
self.fields[11] = [True] * 5
self.fields[23] = [False] * 2
self.fields[18] = [True] * 5
self.fields[16] = [True] * 3
self.fields[12] = [False] * 5
self.pieces = list()
self.piece_size = 42
self.ping_iter = 0
ident = 1
for field_id, field in enumerate(self.fields):
top = field_id // 12 == 1
for piece_id, is_black in enumerate(field):
offset_x = self.board.triangle_width//2 + \
self.board.triangle_width * (field_id % 12) + \
((field_id % 12) // 6) * self.board.offset_x
x = offset_x if top else self.width - offset_x
((field_id % 12) // 6) * self.board.offset_x
y = self.piece_size * \
(piece_id*2+1) if top else self.height - \
self.piece_size * (piece_id*2+1)
pos = (x, y)
self.pieces.append(Piece(self, ident, pos, is_black))
ident += 1
self.dice.reset()
if self.reset_sound is not None:
self.reset_sound.play()
if send:
connection.Send({"action": "resetboard"})
def send_gamestate(self):
pieces = list()
for p in self.pieces:
p.send_move()
self.dice.send_state()
self.dice.send_eyes()
def on_init(self):
pygame.init()
pygame.mixer.init()
self.reset_sound = pygame.mixer.Sound('sound/button.wav')
self.impact_sound = pygame.mixer.Sound('sound/impact.wav')
self.font = pygame.font.Font(pygame.font.get_default_font(), 22)
pygame.display.set_caption('Backgammon')
self.clock = pygame.time.Clock()
self._screen = pygame.display.set_mode(
self.size, pygame.HWSURFACE | pygame.DOUBLEBUF)
self._running = True
def ping(self):
connection.Send({"action": "ping"})
def keep_connection_alive(self):
# Ping every 4 seconds
self.ping_iter = (self.ping_iter + 1) % 240
if self.ping_iter == 0:
self.ping()
def on_event(self, event):
if event.type == pygame.QUIT:
self._running = False
elif event.type == pygame.KEYDOWN:
if event.key == pygame.K_SPACE:
self.dice.roll()
elif event.key == pygame.K_ESCAPE:
self.init_pieces()
else:
self.handle_piece_events(event)
if event.type == pygame.MOUSEMOTION:
connection.Send({'action': 'mousemotion', 'pos': event.pos})
def handle_piece_events(self, event):
for idx, piece in enumerate(self.pieces):
if piece.handle_event(event):
if idx == 0:
break
for idx2, piece2 in enumerate(self.pieces):
if idx == idx2:
continue
if piece.rect.colliderect(piece2.rect):
break
else:
self.pieces.insert(0, self.pieces.pop(idx))
break
else:
self.dice.handle_event(event)
def on_loop(self):
self.keep_connection_alive()
connection.Pump()
self.Pump()
if self.run_server:
self.server.Pump()
def on_render(self):
self.board.render(self._screen)
for piece in self.pieces[::-1]:
piece.update(self._screen)
self.dice.render(self._screen)
self.other_mouse.render(self._screen)
pygame.display.flip()
def on_cleanup(self):
pygame.quit()
def on_execute(self):
if self.on_init() == False:
self._running = False
while (self._running):
self.clock.tick(60)
for event in pygame.event.get():
self.on_event(event)
self.on_loop()
self.on_render()
self.on_cleanup()
def Network_connected(self, data):
print("Connected to the server")
def Network_disconnected(self, data):
print("Disconnected from the server")
self.player_count = 0
def Network_resetboard(self, data):
self.init_pieces(False)
def Network_roll(self, data):
self.dice.roll(data)
def Network_impact(self, data):
self.impact_sound.play()
def Network_eyes(self, data):
self.dice.set_eye_counter(data['eyes'])
def Network_pong(self, data):
pass
def Network_mousemotion(self, data):
self.other_mouse.setPostion(data['pos'])
def Network_playercount(self, data):
new_player_count = int(data['count'])
if self.run_server and new_player_count > self.player_count:
self.send_gamestate()
self.player_count = new_player_count
if self.player_count < 2:
self.other_mouse.set_visible(False)
def Network_move(self, data):
piece_move = data['piece']
for piece in self.pieces:
if piece.ident == piece_move[0]:
piece.move((piece_move[1], piece_move[2]), self._screen)
break
else:
raise ValueError('Invalid piece ident!')
def main():
parser = argparse.ArgumentParser(description='Test')
parser.add_argument('--replay_memory_size',
default=50000,
type=int,
help='replayMemory_size to store training data')
parser.add_argument('--batch_size',
default=512,
type=int,
help='batch size')
parser.add_argument('--learning_rate',
default=1e-3,
type=float,
help='learning_rate')
parser.add_argument('--evaluate_freq',
default=50,
type=int,
help='evaluate once every #evaluate_freq games')
parser.add_argument(
'--train_freq',
default=1,
type=int,
help='train #train_epoch times replay mempry within each train')
parser.add_argument('--n_eval_game',
default=10,
type=int,
help='number of games during one evaluation')
parser.add_argument('--n_burn_in',
default=10,
type=int,
help='number of games to burn in the replay memory')
parser.add_argument('--n_iteration',
default=20,
type=int,
help='number of train iteration')
parser.add_argument('--width', default=6, type=int)
parser.add_argument('--height', default=6, type=int)
parser.add_argument('--n_in_row', default=4, type=int)
args = parser.parse_args()
width, height = args.width, args.height
board = Board(width=width, height=height, n_in_row=args.n_in_row)
game = Game(board)
# Prepare train and eval model
AlphaGoNet_train = PolicyValueNet(width, height)
#AlphaGoNet_best = PolicyValueNet(width, height)
#torch.save(AlphaGoNet_train.policy_value_net.state_dict(), 'model/init.mdl')
AlphaGoNet_train.policy_value_net.load_state_dict(
torch.load('model/current.mdl'))
# Replay is used to store training data:
ReplayMemory = deque(maxlen=args.replay_memory_size)
player = AlphaGoPlayer(NN_fn=AlphaGoNet_train.policy_value_fn)
#eval_player = AlphaGoPlayer(NN_fn=AlphaGoNet_best.policy_value_fn)
eval_player = MCTSPlayer()
max_win_ratio = .0
# Burn in
burn_in(game, player, ReplayMemory, args.n_burn_in)
for i in range(args.n_iteration):
print 'Iteration NO.:', i
train_one_iteration(game, player, ReplayMemory, AlphaGoNet_train,
args.batch_size, args.learning_rate,
args.train_freq, args.evaluate_freq)
win_ratio = evaluate(game, player, eval_player, args.n_eval_game)
if win_ratio > max_win_ratio:
print('Get current_best model!')
max_win_ratio = win_ratio
torch.save(AlphaGoNet_train.policy_value_net.state_dict(),
'model/current_best.mdl')
else:
print('Save current model')
torch.save(AlphaGoNet_train.policy_value_net.state_dict(),
'model/current.mdl')
def __init__(self, init_model=None):
self.writer = SummaryWriter(WRITER_DIR)
# params of the board and the game
self.board_width = 6
self.board_height = 6
self.n_in_row = 4
self.board = Board(width=self.board_width,
height=self.board_height,
n_in_row=self.n_in_row)
self.game = Game(self.board)
# training params
self.learn_rate = 2e-3
self.lr_multiplier = 1.0 # adaptively adjust the learning rate based on KL
self.temp = 1.0 # the temperature param
self.buffer_size = 10000
self.batch_size = 512 # mini-batch size for training
self.data_buffer = deque(maxlen=self.buffer_size)
self.play_batch_size = 1
self.epochs = 5 # num of train_steps for each update
self.kl_targ = 0.02
self.check_freq = 50
self.game_batch_num = 5000
self.improvement_counter = 1000
self.best_win_ratio = 0.0
self.input_plains_num = INPUT_PLANES_NUM
self.c_puct = 5
self.n_playout = 50 # num of simulations for each move
self.shutter_threshold_availables = 1
self.full_boards_selfplay = False
# num of simulations used for the pure mcts, which is used as
# the opponent to evaluate the trained policy
self.pure_mcts_playout_num = 200
self.pure_mcts_playout_num_step = 200
if init_model:
# start training from an initial policy-value net
self.policy_value_net = PolicyValueNet(
self.board_width,
self.board_height,
self.input_plains_num,
model_file=init_model,
shutter_threshold_availables=self.shutter_threshold_availables)
else:
# start training from a new policy-value net
self.policy_value_net = PolicyValueNet(
self.board_width,
self.board_height,
self.input_plains_num,
shutter_threshold_availables=self.shutter_threshold_availables)
self.mcts_player = MCTSPlayer(self.policy_value_net.policy_value_fn,
c_puct=self.c_puct,
n_playout=self.n_playout,
is_selfplay=1)
best_score = float('+inf')
best_move = None
for legal_move in current_board.current_player.calculate_legal_moves():
new_board = legal_move.execute()
move_value, move_obj = minimax(new_board, depth - 1, True)
if move_value < best_score:
best_score = move_value
best_move = legal_move
return best_score, best_move
DEPTH = 2
current_board = Board.create_standard_board()
while True:
_, white_move = minimax(current_board, DEPTH, True)
current_board = white_move.execute()
print(current_board)
print(_)
input()
_, black_move = minimax(current_board, DEPTH, True)
current_board = black_move.execute()
for i in current_board.current_player.calculate_legal_moves():
print(i)
print(current_board)
print(_)
input()
move = -1
if move == -1 or move not in board.availables:
print("invalid move")
move = self.get_action(board)
return move
def __str__(self):
return "Human {}".format(self.player)
n = 5
width, height = 8, 8
model_file = 'best_policy_8_8_5.model'
board = Board(width=width, height=height, n_in_row=n)
game = Game(board)
# ############### human VS AI ###################
# load the trained policy_value_net in either Theano/Lasagne, PyTorch or TensorFlow
# best_policy = PolicyValueNet(width, height, model_file = model_file)
# mcts_player = MCTSPlayer(best_policy.policy_value_fn, c_puct=5, n_playout=400)
# load the provided model (trained in Theano/Lasagne) into a MCTS player written in pure numpy
try:
policy_param = pickle.load(open(model_file, 'rb'))
except:
policy_param = pickle.load(open(model_file, 'rb'),
encoding='bytes') # To support python3
#得到策略
