def __init__(self,
game,
checkpoint_directory,
actor=None,
network_save_interval=100,
rollouts=100,
start_game=0,
replay_save_interval=250,
replay_limit=20000,
minibatch_size=50,
replay_file=None,
test_games=50,
nn_steps=1):
self.game = game
self.checkpoint_directory = checkpoint_directory
self.network_save_interval = network_save_interval
self.mcts = MCTS(game,
simulations=rollouts,
default_policy=self.create_default_policy())
self.game_count = start_game
self.replay_save_interval = replay_save_interval
self.replay_buffer = deque(maxlen=replay_limit)
self.rp_count = 0
self.minibatch_size = minibatch_size
self.test_games = test_games
self.nn_steps = nn_steps
if replay_file == 'auto':
self.replay_file = f'{checkpoint_directory}/replays.txt'
else:
self.replay_file = replay_file
if not os.path.exists(checkpoint_directory):
os.makedirs(checkpoint_directory)
if actor:
self.actor = actor
self.save_actor_to_file()
else:
self.actor = self.load_actor_from_file()
if start_game > 0:
self.actor.load_checkpoint(
f'{checkpoint_directory}/game_{start_game}')
if replay_save_interval > replay_limit:
raise ValueError(
f'replay_save_interval ({replay_save_interval}) must be smaller '
f'than replay_limit ({replay_limit})')
if replay_file is not None and replay_file != 'auto':
try:
self.load_replays()
except FileNotFoundError:
pass
if start_game == 0:
self.actor.save_checkpoint(checkpoint_directory + '/game_0')
self.actor.save_checkpoint(checkpoint_directory + '/best')
with open(checkpoint_directory + '/best.txt', 'w') as f:
f.write(str(0))
class MCTSTrainer:
def __init__(self, gnn, test_graphs, filename):
self.mcts = MCTS(gnn)
self.test_graphs = test_graphs
self.test_result = []
self.filename = filename
# rollout until the end
def train1(self, graph, TAU, batch_size=10, iter_p=2):
self.mcts.train(graph, TAU, batch_size=batch_size, iter_p=iter_p)
# rollout only until leaf
def train2(self, graph, TAU, batch_size=10, iter_p=2):
self.mcts.train(graph,
TAU,
batch_size=batch_size,
stop_at_leaf=True,
iter_p=2)
def test(self):
result = [self.mcts.search(graph) for graph in self.test_graphs]
print(result)
self.test_result.append(result)
def save_test_result(self):
os.makedirs("log", exist_ok=True)
with open("log/{}.pickle".format(self.filename), mode="wb") as f:
pickle.dump(self.test_result, f)
def save_model(self):
os.makedirs("model", exist_ok=True)
torch.save(self.mcts.gnn.state_dict(),
"model/{}.pth".format(self.filename))
class AI(Player):
def __init__(self, name, m):
Player.__init__(self, name)
self.brain = MCTS(m)
def request_input(self, game: Game, state: State):
self.brain.simulate(game, state)
best_idx = self.brain.get_best_child(state)
return game.gen_child_states(state)[best_idx]
def __init__(self, ip_address=None, verbose=True, auto_test=False):
self.series_id = -1
self.starting_player = -1
self.game_count = 0
self.series_count = 0
self.series_game_count = 0
BasicClientActorAbs.__init__(self,
ip_address,
verbose=verbose,
auto_test=auto_test)
trainer = ActorTrainer(self.hex,
'model/1000x500x100-200',
start_game=250)
#self.actor = trainer.actor
self.actor = MCTS(self.hex, simulations=100)
def use_model(t):
gnn, name, graph = t
np.random.seed()
mcts = MCTS(gnn, performance=True)
Timer.start('all')
result = mcts.search_for_exp(graph, time_limit=10 * 60, min_iter_num=100)
print("graph: {}, result: {}".format(name, result))
print("max: ", max(result))
Timer.end('all')
Timer.print()
return max(result)
def choose_mcts(next_move_types, next_moves, last_move_type, last_move, game,
action_mcts, simulation):
#init mcts
if simulation == False:
#要不起不需要mcst
if len(next_moves) == 0:
print("actions", [430])
return "yaobuqi", []
game_copy = copy.deepcopy(game)
game_copy.players[0].model = "mcts"
game_copy.players[1].model = "random"
game_copy.players[2].model = "random"
mcts = MCTS(tree_policy=UCB1(c=1.41),
default_policy=random_terminal_roll_out,
backup=monte_carlo,
game=game_copy)
#state
s = get_state(game_copy.playrecords, player=1)
#action
actions = get_actions(next_moves, game_copy.actions_lookuptable,
game_copy)
#new state
s = combine(s, actions)
begin = time.time()
best_action, win_prob = mcts(s, n=1000)
game.playrecords.win_prob = round(win_prob, 2)
duration = time.time() - begin
print("actions", actions, "best_action", best_action, "win_prob",
win_prob, "time", duration)
if best_action == 429:
return "buyao", []
elif best_action == 430:
return "yaobuqi", []
else:
best_action_id = actions.index(best_action)
return next_move_types[best_action_id], next_moves[best_action_id]
#mcts simulation
else:
if action_mcts == 429:
return "buyao", []
elif action_mcts == 430:
return "yaobuqi", []
else:
return next_move_types[action_mcts], next_moves[action_mcts]
def uct_play_game():
game = NimGame(15)
search_mgr = SearchMgr()
p1, p2 = MCTS(search_mgr).set_root(game), MCTS(search_mgr).set_root(game)
while game.get_actions():
print(str(game))
a1 = p1.uct(game, iters=100)
a2 = p2.uct(game, iters=1000)
if game.player_just_moved == 1:
# play with values for iter_max and verbose = True
# Player 2
a = a2
else:
# Player 1
a = a1
print('Best Action: ' + str(a) + '\n')
game.take_action(a)
p1.update_root(a)
p2.update_root(a)
if game.get_result(game.player_just_moved) == 1.0:
print('Player ' + str(game.player_just_moved) + ' wins!')
elif game.get_result(game.player_just_moved) == 0.0:
print('Player ' + str(3 - game.player_just_moved) + ' wins!')
else:
print('Nobody wins!')
def play_series(x):
game = Hex()
actor = Actor(game, [], replay_file='model/replays_expert.txt', rp_save_interval=replay_save_interval)
mcts = MCTS(game, simulations=rollouts)
for i in range(games_per_series):
print('Starting game 1')
state = game.get_initial_state()
mcts.set_state(state)
while not game.is_finished(state):
move, probabilities = mcts.select_move(True)
padded_probs = np.pad(probabilities, (0, game.num_possible_moves() - len(probabilities)), 'constant')
actor.add_to_replay_buffer(state, padded_probs)
state = game.get_outcome_state(state, move)
mcts.set_state(state)
params["n_input_features"] = numpy.prod(env.observation_space.shape)
params["env"] = env
params["gamma"] = 0.99
# Planning/MCTS Hyperparameters
params["horizon"] = 10
params["simulations"] = 100 # 1000
# Deep RL Hyperparameters
params["alpha"] = 0.0005 # 0.001
params["epsilon"] = 0.1
params["memory_capacity"] = 10000
params["warmup_phase"] = 1000
params["target_update_interval"] = 5000
params["minibatch_size"] = 64
params["epsilon_linear_decay"] = 1.0 / params["memory_capacity"]
params["epsilon_min"] = 0.01
training_episodes = 1 # 2000
mcts_agent = MCTS(params["env"],
params["gamma"],
c=1.,
n_iter=params["simulations"])
a2c_agent = a2c.A2CLearner(params)
lens = [len(episode(env, mcts_agent, a2c_agent, i)) for i in range(500)]
actions = episode(env, mcts_agent, a2c_agent, 500)
print('-')
print(actions)
plot.plot(lens)
plot.show()
pos = new_chess_game()
for i in range(len(move)):
movei = move[i]
if pos[movei[0],movei[1]] != 0:
pos[movei[2],movei[3]] = pos[movei[0],movei[1]]
pos[movei[0],movei[1]] = 0
else:
ValueErr("error")
print(np.flipud(pos))
mcts = MCTS(tree_policy=Go(c=5),
default_policy=RandomKStepRollOut_Value(20, 0.95),
backup=monte_carlo)
policy_fun = policy_nn()
rollout_fun = rollout_nn()
value_fun = value_nn()
root = StateNode(None, ChessState(pos, 1, policy_fun, rollout_fun, value_fun, False ))
best_action = mcts(root, n=500)
pr.disable()
s = io.StringIO()
sortby = 'cumulative'
ps = pstats.Stats(pr, stream=s).sort_stats(sortby)
ps.print_stats()
print(s.getvalue())
from games.hex import Hex
from games.random_player import RandomPlayer
from mcts.mcts import MCTS
game = Hex()
p1 = MCTS(game, simulations=50)
p2 = RandomPlayer(game)
players = (p1, p2)
games = 20
p2_starting = False
wins = 0
for i in range(games):
state = game.get_initial_state()
turn = p2_starting
while not game.is_finished(state):
for p in players:
p.set_state(state)
move = players[int(turn)].select_move()
state = game.get_outcome_state(state, move)
turn = not turn
result = game.evaluate_state(state)
if p2_starting and result == -1 or not p2_starting and result == 1:
wins += 1
print(f'Won game {i+1}')
else:
print(f'Lost game {i+1}')
p2_starting = not p2_starting
class ActorTrainer:
def __init__(self,
game,
checkpoint_directory,
actor=None,
network_save_interval=100,
rollouts=100,
start_game=0,
replay_save_interval=250,
replay_limit=20000,
minibatch_size=50,
replay_file=None,
test_games=50,
nn_steps=1):
self.game = game
self.checkpoint_directory = checkpoint_directory
self.network_save_interval = network_save_interval
self.mcts = MCTS(game,
simulations=rollouts,
default_policy=self.create_default_policy())
self.game_count = start_game
self.replay_save_interval = replay_save_interval
self.replay_buffer = deque(maxlen=replay_limit)
self.rp_count = 0
self.minibatch_size = minibatch_size
self.test_games = test_games
self.nn_steps = nn_steps
if replay_file == 'auto':
self.replay_file = f'{checkpoint_directory}/replays.txt'
else:
self.replay_file = replay_file
if not os.path.exists(checkpoint_directory):
os.makedirs(checkpoint_directory)
if actor:
self.actor = actor
self.save_actor_to_file()
else:
self.actor = self.load_actor_from_file()
if start_game > 0:
self.actor.load_checkpoint(
f'{checkpoint_directory}/game_{start_game}')
if replay_save_interval > replay_limit:
raise ValueError(
f'replay_save_interval ({replay_save_interval}) must be smaller '
f'than replay_limit ({replay_limit})')
if replay_file is not None and replay_file != 'auto':
try:
self.load_replays()
except FileNotFoundError:
pass
if start_game == 0:
self.actor.save_checkpoint(checkpoint_directory + '/game_0')
self.actor.save_checkpoint(checkpoint_directory + '/best')
with open(checkpoint_directory + '/best.txt', 'w') as f:
f.write(str(0))
def train(self, num_games):
for i in range(num_games):
self.game_count += 1
game_start_time = time.time()
print(f'[GAME {self.game_count}] Initializing state')
state = self.game.get_initial_state()
self.mcts.set_state(state)
print(f'[GAME {self.game_count}] Simulating game')
while not self.game.is_finished(state):
move, probabilities = self.mcts.select_move(True)
padded_probs = np.pad(
probabilities,
(0, self.game.num_possible_moves() - len(probabilities)),
'constant')
self.add_to_replay_buffer(state, padded_probs)
state = game.get_outcome_state(state, move)
self.mcts.set_state(state)
print(f'[GAME {self.game_count}] Training neural network')
for j in range(self.nn_steps):
self.train_network()
if self.game_count % self.network_save_interval == 0:
print(
f'[GAME {self.game_count}] Saving neural network checkpoint'
)
self.actor.save_checkpoint(
f'{self.checkpoint_directory}/game_{self.game_count}')
if self.test_against_best():
print(
f'[GAME {self.game_count}] New best found - saving checkpoint'
)
print(
f'[GAME {self.game_count}] Time elapsed: {time.time() - game_start_time:.2f}'
)
print()
def test_against_best(self):
if self.test_games <= 0:
return False
print(f'[GAME {self.game_count}] Testing against best model...',
end='')
best_actor = self.load_actor_from_file()
best_actor.load_checkpoint(f'{self.checkpoint_directory}/best')
starting = True
wins = 0
for i in range(self.test_games):
turn = starting
state = self.game.get_initial_state()
while not self.game.is_finished(state):
if turn:
move = self.actor.select_move(state)
else:
move = best_actor.select_move(state)
state = game.get_outcome_state(state, move[0])
turn = not turn
result = game.evaluate_state(state)
if result == 1 and starting or result == -1 and not starting:
wins += 1
starting = not starting
print(f'won {wins}/{self.test_games}')
if wins > self.test_games / 2:
self.actor.save_checkpoint(self.checkpoint_directory + '/best')
with open(self.checkpoint_directory + '/best.txt', 'w') as f:
f.write(str(self.game_count))
return True
return False
def train_network(self):
minibatch = random.sample(
self.replay_buffer,
min(self.minibatch_size, len(self.replay_buffer)))
for i in range(len(minibatch)):
minibatch[i] = self.game.format_for_nn(
minibatch[i][0], format=self.actor.format), minibatch[i][1]
self.actor.network.train(minibatch=minibatch)
def create_default_policy(self):
def actor_default_policy(state, moves):
move = self.actor.select_move(state, stochastic=True)
return move
return actor_default_policy
def add_to_replay_buffer(self, state, probabilities):
self.replay_buffer.append((state, probabilities))
self.rp_count += 1
if self.replay_save_interval != -1 and self.rp_count % self.replay_save_interval == 0 and self.rp_count != 0:
replays = len(self.replay_buffer)
self.save_replays(
itertools.islice(self.replay_buffer,
replays - self.replay_save_interval, replays))
def save_replays(self, replays):
if self.replay_file is None:
return
with open(self.replay_file, 'a') as f:
for replay in replays:
state_string = ','.join(map(str, replay[0][0])) + ',' + str(
replay[0][1])
probs_string = ','.join(map(str, replay[1]))
rp_string = state_string + ';' + probs_string
f.write(rp_string + '\n')
def load_replays(self):
with open(self.replay_file, 'r') as f:
for line in f:
state, probs = line.split(';')
state = list(map(int, state.split(',')))
player = state[-1]
board = state[:-1]
probs = list(map(float, probs.split(',')))
self.replay_buffer.append(((board, player), probs))
def load_actor_from_file(self):
with open(f'{self.checkpoint_directory}/actor_params.txt') as f:
lines = f.read().split('\n')
format = lines[0]
optimizer = 'adam'
if len(lines) > 1:
optimizer = lines[1]
with open(f'{self.checkpoint_directory}/actor_layers.bin', 'rb') as f:
layers = pickle.load(f)
return Actor(self.game, layers, format=format, optimizer=optimizer)
def save_actor_to_file(self):
with open(f'{self.checkpoint_directory}/actor_params.txt', 'w') as f:
f.write(self.actor.format + '\n')
f.write(self.actor.optimizer)
with open(f'{self.checkpoint_directory}/actor_layers.bin', 'wb') as f:
pickle.dump(self.actor.layers, f)
def __init__(self, name, m):
Player.__init__(self, name)
self.brain = MCTS(m)
from mcts.mcts import MCTS
from games.nim import Nim
from games.random_player import RandomPlayer
import random
# Instantiate our game with given parameters
game = Nim(9, 3)
num_games = 50
play_mode = 0
# Create a new MCTS player for player 1
player1 = MCTS(game, simulations=1000)
# Create player 2 - either as the same player as player 1, or as a random player
player2 = (player1, RandomPlayer(game))[0]
players = [player1, player2]
def run_single_game(starting_player=0, verbose=False):
"""
Runs a simulation of a single game, and returns the winning player.
:param starting_player: The player that should start the game.
:param verbose: If True, string representations of all moves will be printed to the console.
:return: 0 if player 1 is the winner, 1 if player 2 is the winner.
"""
state = game.get_initial_state(starting_player)
current_player = starting_player
for p in players:
p.set_state(state)
while not game.is_finished(state):
move = players[current_player].select_move()
class BasicClientActor(BasicClientActorAbs):
def __init__(self, ip_address=None, verbose=True, auto_test=False):
self.series_id = -1
self.starting_player = -1
self.game_count = 0
self.series_count = 0
self.series_game_count = 0
BasicClientActorAbs.__init__(self,
ip_address,
verbose=verbose,
auto_test=auto_test)
trainer = ActorTrainer(self.hex,
'model/1000x500x100-200',
start_game=250)
#self.actor = trainer.actor
self.actor = MCTS(self.hex, simulations=100)
def handle_get_action(self, state):
"""
Here you will use the neural net that you trained using MCTS to select a move for your actor on the current
board. Remember to use the correct player_number for YOUR actor! The default action is to select a random empty
cell on the board. This should be modified.
:param state: The current board in the form (1 or 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0), where
1 or 2 indicates the number of the current player. If you are player 2 in the current series, for example,
then you will see a 2 here throughout the entire series, whereas player 1 will see a 1.
:return: Your actor's selected action as a tuple (row, column)
"""
current_player = state[0] - 1
board = list(state[1:])
state = (board, current_player)
#next_move = self.actor.select_move(state)[0][0]
self.actor.set_state(state)
next_move = self.actor.select_move()[0]
return next_move
def handle_series_start(self, unique_id, series_id, player_map, num_games,
game_params):
"""
Set the player_number of our actor, so that we can tell our MCTS which actor we are.
:param unique_id: integer identifier for the player within the whole tournament database
:param series_id: (1 or 2) indicating which player this will be for the ENTIRE series
:param player_map: a list of tuples: (unique-id series-id) for all players in a series
:param num_games: number of games to be played in the series
:param game_params: important game parameters. For Hex = list with one item = board size (e.g. 5)
:return
"""
self.series_id = series_id
self.series_count += 1
print(f'Series {self.series_count} starting')
print(f'Series ID: {series_id}')
self.series_game_count = 0
#############################
#
#
# YOUR CODE (if you have anything else) HERE
#
#
##############################
def handle_game_start(self, start_player):
"""
:param start_player: The starting player number (1 or 2) for this particular game.
:return
"""
self.starting_player = start_player
self.game_count += 1
print(
f'Game {self.game_count} starting. (Game {self.series_game_count} in series.)'
)
#############################
#
#
# YOUR CODE (if you have anything else) HERE
#
#
##############################
def handle_game_over(self, winner, end_state):
"""
Here you can decide how to handle what happens when a game finishes. The default action is to print the winner
and the end state.
:param winner: Winner ID (1 or 2)
:param end_state: Final state of the board.
:return:
"""
#############################
#
#
# YOUR CODE HERE
#
#
##############################
print()
print("Game over, these are the stats:")
print('Winner: ' + str(winner))
print('End state:')
self.print_state(end_state)
def handle_series_over(self, stats):
"""
Here you can handle the series end in any way you want; the initial handling just prints the stats.
:param stats: The actor statistics for a series = list of tuples [(unique_id, series_id, wins, losses)...]
:return:
"""
#############################
#
#
# YOUR CODE HERE
#
#
#############################
print("Series ended, these are the stats:")
print(f'Series ID: {self.series_id}')
for stat in stats:
if stat[1] == self.series_id:
# Found my stats
print(
f'Won {stat[2]}/{stat[2] + stat[3]} ({stat[2]/(stat[2]+stat[3]):.0%})'
)
print()
# print(str(stats))
def handle_tournament_over(self, score):
"""
Here you can decide to do something when a tournament ends. The default action is to print the received score.
:param score: The actor score for the tournament
:return:
"""
#############################
#
#
# YOUR CODE HERE
#
#
#############################
print("Tournament over. Your score was: " + str(score))
def handle_illegal_action(self, state, illegal_action):
"""
Here you can handle what happens if you get an illegal action message. The default is to print the state and the
illegal action.
:param state: The state
:param illegal_action: The illegal action
:return:
"""
#############################
#
#
# YOUR CODE HERE
#
#
#############################
print("An illegal action was attempted:")
print('State: ' + str(state))
print('Action: ' + str(illegal_action))
def __init__(self, gnn, test_graphs, filename):
self.mcts = MCTS(gnn)
self.test_graphs = test_graphs
self.test_result = []
self.filename = filename
testfile = "trivialExample.in"
problem = FileReader(testfile)
print("Video Sizes: %r" % (problem.videoSizes,))
print("Endpoints:\n\t%s" % ("\n\t".join([str(e) for e in problem.endpoints])))
print("Requests: %r" % ([r for r in problem.requests]))
# Generate initial state
initial_contents = list([(0, []) for _ in range(problem.nCaches)])
initial_score = 0
initial_state = TreeState(caches_contents=initial_contents, score=initial_score,
problem=problem)
# Generate the optimal end state
mcts = MCTS(tree_policy=UCB1(c=1.41),
default_policy=immediate_reward,
backup=monte_carlo)
node = StateNode(parent=None, state=initial_state)
while True:
if node.state.is_terminal():
print("Terminal node reached.")
break
print("Finding best action")
best_action = mcts(node)
print("Performing action")
node = StateNode(parent=None, state=node.state.perform(best_action))
print("Score now is: %d" % node.state.score)
print("Saving output")
def mcts_agent(game: ".game.Game", move_number) -> ".game.Game":
mcts = MCTS(game, move_number)
cards_to_choose, cards_attack = mcts.choose_next_move()
choose_card_from_hand(game, ChooseCard.DEFINED_CARDS, cards_to_choose)
attack_opponent(game, AttackOpponent.DEFINED_CARDS, cards_attack)
timeout = int(args.timeout) - 5
c1 = Client(args.ip, PORTS[player_arg], player_arg)
c1.send_name(PLAYER_NAMES[player_arg])
board = Board()
game = Game(board)
# Main game loop
try:
while not game.ended:
state = None
while state is None:
state, turn = c1.receive_state()
game.board.board = state
if turn not in ["black", "white"]:
raise GameEndedException
game.turn = TURN_MAPPING[turn]
print(state, turn)
if game.turn == OUR_PLAYER:
mcts = MCTS(deepcopy(game),
OUR_PLAYER,
max_depth=max_depth,
C=C)
start, end = mcts.search(timeout)
print(start, end)
c1.send_move(start, end)
except GameEndedException:
print("Game ended with state {}".format(turn))
c1.close()
