def evaluate_state(self, env: MancalaEnv) -> (float, float):
flip_board = env.side_to_move == Side.NORTH
state = env.board.get_board_image(flipped=flip_board)
mask = env.get_action_mask_with_no_pie()
dist, _, value = self.network.evaluate_move(state=state, mask=mask)
return dist, float(value)
def mcts_main():
mcts = TreesFactory.standard_mcts()
state = MancalaEnv()
try:
_run_game(mcts, state)
except Exception as e:
logging.error("Uncaught exception in main: " + str(e))
def test_mcts_doesnt_mutate_state(self):
state = MancalaEnv()
initial_board = Board.clone(state.board)
mcts = MCTSFactory.test_mcts()
mcts.search(state)
self.assertEqual(initial_board.board, state.board.board, "Expect MCTS doesn't mutate the initial board")
def start():
_state = MancalaEnv()
_player = RandomPlayer()
try:
_run_game(_player, _state)
except Exception as e:
print("Uncaught exception in main: " + str(e))
def __init__(self, state: MancalaEnv, move: Move = None, parent=None):
self.visits = 0
self.reward = 0
self.state = state
self.children = []
self.parent = parent
self.move = move
self.value = -1
self.unexplored_moves = set(state.get_legal_moves())
def main(_):
with tf.Session() as sess:
with tf.variable_scope("global"):
a3client = A3Client(sess)
mcts = TreesFactory.alpha_mcts(a3client)
state = MancalaEnv()
try:
_run_game(mcts, state)
except Exception as e:
logging.error("Uncaught exception in main: " + str(e))
def search(self, state: MancalaEnv) -> Move:
# short circuit last move
if len(state.get_legal_moves()) == 1:
return state.get_legal_moves()[0]
game_state_root = Node(state=MancalaEnv.clone(state))
start_time = datetime.datetime.utcnow()
games_played = 0
while datetime.datetime.utcnow() - start_time < self.calculation_time:
node = self.tree_policy.select(game_state_root)
final_state = self.default_policy.simulate(node)
self.rollout_policy.backpropagate(node, final_state)
# Debugging information
games_played += 1
logging.debug("%s; Game played %i" % (node, games_played))
logging.debug("%s" % game_state_root)
chosen_child = node_utils.select_robust_child(game_state_root)
logging.info("Choosing: %s" % chosen_child)
return chosen_child.move
def expand(parent: Node) -> Node:
child_expansion_move = choice(tuple(parent.unexplored_moves))
child_state = MancalaEnv.clone(parent.state)
child_state.perform_move(child_expansion_move)
child_node = Node(state=child_state,
move=child_expansion_move,
parent=parent)
parent.put_child(child_node)
MonteCarloTreePolicy._rave_expand(child_node)
# go down the tree
return child_node
def _alpha_beta_search(game: MancalaEnv, alpha=-np.inf, beta=np.inf, depth=5):
"""Search game to determine best action; use alpha-beta pruning.
This version cuts off search and uses an evaluation function."""
if depth == 0 or game.is_game_over():
return game.get_player_utility()
if game.side_to_move == Side.SOUTH:
v = -np.inf
for (_, new_s) in game.next_states():
v = max(v, AlphaBeta._alpha_beta_search(new_s, alpha, beta, depth - 1))
alpha = max(alpha, v)
# if beta <= alpha:
# break
else:
v = np.inf
for (_, new_s) in game.next_states():
v = min(v, AlphaBeta._alpha_beta_search(new_s, alpha, beta, depth - 1))
beta = min(beta, v)
# if beta <= alpha:
# break
return v
def _rave_expand(parent: Node):
moves = [-1e80 for _ in range(parent.state.board.holes + 1)]
for unexplored_move in parent.unexplored_moves.copy():
child_state = MancalaEnv.clone(parent.state)
child_state.perform_move(unexplored_move)
moves[unexplored_move.index] = evaluation.get_score(
state=child_state, parent_side=parent.state.side_to_move)
moves_dist = np.asarray(moves, dtype=np.float64).flatten()
exp = np.exp(moves_dist - np.max(moves_dist))
dist = exp / np.sum(exp)
parent.value = max(dist)
def backpropagate(self, root: Node, final_state: MancalaEnv):
"""
backpropgate pushes the reward (pay/visits) to the parents node up to the root
:param root: starting node to backpropgate from
:param final_state: the state of final node (holds final reward from the simulation)
"""
node = root
# propagate node reward to parents'
while node is not None:
side = node.parent.state.side_to_move if node.parent is not None else node.state.side_to_move # root node
node.update(final_state.compute_end_game_reward(side))
node = node.parent
def expand(self, node: AlphaNode):
# Tactical workaround the pie move
if Move(node.state.side_to_move, 0) in node.unexplored_moves:
node.unexplored_moves.remove(Move(node.state.side_to_move, 0))
dist, value = self.network.evaluate_state(node.state)
for index, prior in enumerate(dist):
expansion_move = Move(node.state.side_to_move, index + 1)
if node.state.is_legal(expansion_move):
child_state = MancalaEnv.clone(node.state)
child_state.perform_move(expansion_move)
child_node = AlphaNode(state=child_state,
prior=prior,
move=expansion_move,
parent=node)
node.put_child(child_node)
# go down the tree
return node_utils.select_child_with_maximum_action_value(node)
def backpropagate(self, root: Node, final_state: MancalaEnv, lmbd=1):
"""backpropgate pushes the reward (pay/visits) to the parents node starting from the root down
:param root: starting node to backpropgate from
:param final_state: the state of final node (holds final reward from the simulation)
:param lmbd: a parameter to control the weight of the value network
"""
path_stack = []
node = root
# propagate node reward to parents'
while node is not None:
path_stack.append(node)
node = node.parent
# Update from root downward so the exploration bonus calculation is correct
while len(path_stack) > 0:
node = path_stack.pop()
side = node.parent.state.side_to_move if node.parent is not None else node.state.side_to_move # root node
game_reward = final_state.compute_end_game_reward(side)
# _, value = self.network.evaluate_state(final_state)
# game_reward = (1 - lmbd) * value + (lmbd * side_final_reward) # value from network + value from actionNet
node.update(game_reward)
def _run_game(player: Player, state: MancalaEnv):
our_agent_states = []
their_agent_states = []
both_agent_states = []
our_side = Side.SOUTH
while True:
msg = protocol.read_msg()
try:
msg_type = protocol.get_msg_type(msg)
if msg_type == MsgType.START:
first = protocol.interpret_start_msg(msg)
if first:
move = player.get_play(state)
protocol.send_msg(protocol.create_move_msg(move.index))
else:
our_side = Side.NORTH
elif msg_type == MsgType.STATE:
move_turn = protocol.interpret_state_msg(msg)
if move_turn.move == 0:
our_side = Side.opposite(our_side)
move_to_perform = Move(state.side_to_move, move_turn.move)
observed_state = ObservedState(state=state,
action_taken=move_to_perform)
both_agent_states.append(observed_state)
if state.side_to_move == our_side:
our_agent_states.append(observed_state)
else:
their_agent_states.append(observed_state)
state.perform_move(move_to_perform)
if not move_turn.end:
if move_turn.again:
move = player.get_play(state)
# pie rule; optimal move is to swap
if move.index == 0:
protocol.send_msg(protocol.create_swap_msg())
else:
protocol.send_msg(
protocol.create_move_msg(move.index))
elif msg_type == MsgType.END:
args = parser.parse_args()
run_id = '%06d' % args.run_number
run_category = args.category
_our_agent_file_path = _checkpoint_file_path + "/our-agent/" + run_category + run_id
_their_agent_file_path = _checkpoint_file_path + "/their-agent/" + run_category + run_id
_both_agent_file_path = _checkpoint_file_path + "/both-agent/" + run_category + run_id
np.save(file=_our_agent_file_path,
arr=np.array(our_agent_states))
np.save(file=_their_agent_file_path,
arr=np.array(their_agent_states))
np.save(file=_both_agent_file_path,
arr=np.array(both_agent_states))
break
else:
print("Not sure what I got " + str(msg_type))
except InvalidMessageException as _e:
print(str(_e))
def test_is_legal_move_returns_true_for_the_pie_rule(self):
board = self.game.board
MancalaEnv.make_move(board, Move(Side.SOUTH, 6), False)
self.assertTrue(MancalaEnv.is_legal_action(board, Move(Side.NORTH, 0), False))
def setUp(self):
self.game = MancalaEnv()
def test_mcts_test_game(self):
state = MancalaEnv()
mcts = MCTSFactory.long_test_mcts(sec=0) # Tweak this to test MCTS manually
move = mcts.search(state)
print(move)
def test_mcts_generate_legal_move(self):
state = MancalaEnv()
mcts = MCTSFactory.test_mcts()
move = mcts.search(state)
self.assertTrue(state.is_legal(move), "Expect move generated by MCTS is legal move")
def sample_state(self, env: MancalaEnv) -> (int, float):
flip_board = env.side_to_move == Side.NORTH
state = env.board.get_board_image(flipped=flip_board)
mask = env.get_action_mask_with_no_pie()
return self.network.sample(state=state, mask=mask)
def test_is_legal_move_returns_true_for_the_pie_rule2(self):
env = MancalaEnv()
env.perform_move(Move(Side.SOUTH, 5))
self.assertTrue(env.is_legal(Move(Side.NORTH, 0)))
def search(self, game: MancalaEnv) -> Move:
values = [(a, self._alpha_beta_search(game=state, depth=self.depth)) for a, state in game.next_states()]
np.random.shuffle(values)
if game.side_to_move == Side.SOUTH:
action, _ = max(values, key=lambda x: x[1])
else:
action, _ = min(values, key=lambda x: x[1])
return action
def run(args, server):
env = MancalaEnv()
trainer = A3C(env, args.task)
# Variable names that start with "local" are not saved in checkpoints.
variables_to_save = [
v for v in tf.global_variables() if not v.name.startswith("local")
]
init_op = tf.variables_initializer(variables_to_save)
init_all_op = tf.global_variables_initializer()
saver = FastSaver(variables_to_save)
var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
tf.get_variable_scope().name)
logger.info('Trainable vars:')
for v in var_list:
logger.info(' %s %s', v.name, v.get_shape())
def init_fn(ses):
logger.info("Initializing all parameters.")
ses.run(init_all_op)
config = tf.ConfigProto(device_filters=[
"/job:ps", "/job:worker/task:{}/cpu:0".format(args.task)
])
logdir = os.path.join(args.log_dir, 'train')
summary_writer = tf.summary.FileWriter(logdir + "_%d" % args.task)
logger.info("Events directory: %s_%s", logdir, args.task)
sv = tf.train.Supervisor(
is_chief=(args.task == 0),
logdir=logdir,
saver=saver,
summary_op=None,
init_op=init_op,
init_fn=init_fn,
summary_writer=summary_writer,
ready_op=tf.report_uninitialized_variables(variables_to_save),
global_step=trainer.global_step,
save_model_secs=30,
save_summaries_secs=30)
num_global_steps = 100000000
logger.info(
"Starting session. If this hangs, we're mostly likely waiting to connect to the parameter server. "
+
"One common cause is that the parameter server DNS name isn't resolving yet, or is misspecified."
)
with sv.managed_session(server.target,
config=config) as sess, sess.as_default():
sess.run(trainer.down_sync)
global_step = sess.run(trainer.global_step)
logger.info("Starting training at step=%d", global_step)
while not sv.should_stop() and (not num_global_steps
or global_step < num_global_steps):
trainer.play(sess, RandomAgent(), summary_writer)
global_step = sess.run(trainer.global_step)
# Ask for all the services to stop.
sv.stop()
logger.info('reached %s steps. worker stopped.', global_step)
class TestMancalaGameState(unittest.TestCase):
def setUp(self):
self.game = MancalaEnv()
def test_initial_state_is_correct(self):
self.assertEqual(self.game.side_to_move, Side.SOUTH)
self.assertFalse(self.game.north_moved)
for hole in range(1, self.game.board.holes + 1):
self.assertEqual(self.game.board.get_seeds(Side.SOUTH, hole), 7)
self.assertEqual(self.game.board.get_seeds(Side.NORTH, hole), 7)
self.assertEqual(self.game.board.get_seeds_in_store(Side.SOUTH), 0)
self.assertEqual(self.game.board.get_seeds_in_store(Side.NORTH), 0)
def test_cloning_immutability(self):
clone = MancalaEnv.clone(self.game)
self.game.perform_move(Move(Side.SOUTH, 3))
self.assertEqual(clone.board.get_seeds(Side.SOUTH, 3), 7)
self.assertEqual(clone.side_to_move, Side.SOUTH)
def test_move_has_required_effects(self):
self.game.perform_move(Move(Side.SOUTH, 5))
self.assertEqual(self.game.board.get_seeds(Side.SOUTH, 5), 0)
self.assertEqual(self.game.board.get_seeds(Side.SOUTH, 6), 8)
self.assertEqual(self.game.board.get_seeds(Side.SOUTH, 7), 8)
self.assertEqual(self.game.board.get_seeds_in_store(Side.SOUTH), 1)
self.assertEqual(self.game.board.get_seeds(Side.NORTH, 1), 8)
self.assertEqual(self.game.board.get_seeds(Side.NORTH, 2), 8)
self.assertEqual(self.game.board.get_seeds(Side.NORTH, 3), 8)
self.assertEqual(self.game.board.get_seeds(Side.NORTH, 4), 8)
self.game.perform_move(Move(Side.NORTH, 4))
self.assertEqual(self.game.board.get_seeds(Side.NORTH, 4), 0)
self.assertEqual(self.game.board.get_seeds(Side.NORTH, 5), 8)
self.assertEqual(self.game.board.get_seeds(Side.NORTH, 6), 8)
self.assertEqual(self.game.board.get_seeds(Side.NORTH, 7), 8)
self.assertEqual(self.game.board.get_seeds_in_store(Side.NORTH), 1)
self.assertEqual(self.game.board.get_seeds(Side.SOUTH, 1), 8)
self.assertEqual(self.game.board.get_seeds(Side.SOUTH, 2), 8)
self.assertEqual(self.game.board.get_seeds(Side.SOUTH, 3), 8)
def test_game_is_over_returns_false(self):
self.assertFalse(self.game.is_game_over())
def test_game_is_over_returns_true(self):
board = self.game.board
for hole in range(1, board.holes + 1):
board.set_seeds(Side.SOUTH, hole, 0)
board.set_seeds_in_store(Side.SOUTH, 49)
self.assertTrue(self.game.is_game_over())
def test_game_returns_winner_the_player_with_most_seeds(self):
board = self.game.board
for hole in range(1, self.game.board.holes + 1):
board.set_seeds(Side.SOUTH, hole, 0)
board.set_seeds(Side.NORTH, hole, 0)
board.set_seeds_in_store(Side.SOUTH, 23)
board.set_seeds_in_store(Side.NORTH, 21)
self.assertEqual(self.game.get_winner(), Side.SOUTH)
def test_game_returns_no_winner_if_players_have_equal_number_of_seeds(self):
board = self.game.board
for hole in range(1, self.game.board.holes + 1):
board.set_seeds(Side.SOUTH, hole, 0)
board.set_seeds(Side.NORTH, hole, 0)
board.set_seeds_in_store(Side.SOUTH, 30)
board.set_seeds_in_store(Side.NORTH, 30)
self.assertEqual(self.game.get_winner(), None)
def test_is_legal_move_returns_true_for_the_pie_rule(self):
board = self.game.board
MancalaEnv.make_move(board, Move(Side.SOUTH, 6), False)
self.assertTrue(MancalaEnv.is_legal_action(board, Move(Side.NORTH, 0), False))
def test_is_legal_move_returns_true_for_the_pie_rule2(self):
env = MancalaEnv()
env.perform_move(Move(Side.SOUTH, 5))
self.assertTrue(env.is_legal(Move(Side.NORTH, 0)))
def test_legal_moves_contains_all_moves(self):
self.assertEqual(len(set(self.game.get_legal_moves())), 7)
self.game.perform_move(Move(Side.SOUTH, 3))
self.assertEqual(len(set(self.game.get_legal_moves())), 8)
def test_side_to_move_doesnt_change(self):
self.game.perform_move(Move(Side.SOUTH, 1))
self.assertEqual(self.game.side_to_move, Side.NORTH)
def test_alphbeta(self):
board = self.game.board
for hole in range(1, self.game.board.holes + 1):
board.set_seeds(Side.SOUTH, hole, 0)
board.set_seeds(Side.NORTH, hole, 0)
board.set_seeds_in_store(Side.SOUTH, 0)
board.set_seeds_in_store(Side.NORTH, 0)
board.set_seeds(Side.SOUTH, 3, 1)
board.set_seeds(Side.SOUTH, 2, 4)
board.set_seeds_op(Side.SOUTH, 4, 5)
print(self.game)
print(search_action(self.game))
def __init__(self, env: MancalaEnv, task: int):
self.env = env
self.task = task
# Performance statistics
self.episodes_reward = []
self.episodes_length = []
self.episodes_mean_value = []
self.wins = 0
self.games = 0
worker_device = "/job:worker/task:{}/cpu:0".format(task)
with tf.device(
tf.train.replica_device_setter(1,
worker_device=worker_device)):
with tf.variable_scope("global"):
# The input board is a tensor 2 x 8 x 1. The last dimension is added so that
# convolutional layers can be applied to the input
self.network = ACNetwork(state_shape=[2, 8, 1], num_act=7)
self.global_step = tf.get_variable(
"global_step", [],
tf.int32,
initializer=tf.constant_initializer(0, dtype=tf.int32),
trainable=False)
with tf.device(worker_device):
with tf.variable_scope("local"):
self.local_network = pi = self.network
pi.global_step = self.global_step
self.action = tf.placeholder(shape=[None], dtype=tf.int32)
self.action_one_hot = tf.one_hot(self.action, 7, dtype=tf.float32)
self.target_v = tf.placeholder(shape=[None], dtype=tf.float32)
self.advantage = tf.placeholder(shape=[None], dtype=tf.float32)
log_prob = tf.nn.log_softmax(pi.logits)
prob = tf.nn.softmax(pi.logits)
act_log_prob = tf.reduce_sum(log_prob * self.action_one_hot, [1])
# Loss functions
self.value_loss = 0.5 * tf.reduce_sum(
tf.square(self.target_v - tf.reshape(pi.value, [-1])))
self.entropy = -tf.reduce_sum(prob * log_prob)
self.policy_loss = -tf.reduce_sum(act_log_prob * self.advantage)
# self.reg_loss = tf.add_n([tf.nn.l2_loss(v) for v in local_vars])
self.loss = 0.5 * self.value_loss + self.policy_loss - self.entropy * 0.01 # + 0.002 * self.reg_loss
# Get gradients from local network using local losses and clip them to avoid exploding gradients
self.gradients = tf.gradients(self.loss, pi.vars)
grads, self.grad_norms = tf.clip_by_global_norm(
self.gradients, 100.0)
# Define operation for downloading the weights from the parameter server (ps)
# on the local model of the worker
self.down_sync = tf.group(
*[v1.assign(v2) for v1, v2 in zip(pi.vars, self.network.vars)])
# Define the training operation which applies the gradients on the parameter server network (up sync)
optimiser = tf.train.RMSPropOptimizer(learning_rate=0.0007)
grads_and_global_vars = list(zip(grads, self.network.vars))
inc_step = self.global_step.assign_add(tf.shape(self.action)[0])
self.train_op = tf.group(
*[optimiser.apply_gradients(grads_and_global_vars), inc_step])
# Define an environment runner of this network
self.env_runner = EnvironmentRunner(MancalaEnv(), pi)
episode_size = tf.to_float(tf.shape(pi.value)[0])
# Define summaries for tensorboard
tf.summary.scalar("Model/PolicyLoss",
self.policy_loss / episode_size)
tf.summary.scalar("Model/ValueLoss",
self.value_loss / episode_size)
tf.summary.scalar("Model/Entropy", self.entropy / episode_size)
tf.summary.scalar("Model/GradientsGlobalNorm", self.grad_norms)
tf.summary.scalar("Model/VarGlobalNorm", tf.global_norm(pi.vars))
self.summary_op = tf.summary.merge_all()
self.summary_writer = None
self.local_steps = 0
def __init__(self, state: MancalaEnv, action_taken: Move):
self.state = MancalaEnv.clone(state)
self.action_taken = Move.clone(action_taken)
def get_play(state: MancalaEnv) -> Move:
return choice(state.get_legal_moves())
def _make_temp_child(parent: Node, move: Move) -> MancalaEnv:
child_state = MancalaEnv.clone(parent.state)
child_state.perform_move(move)
return child_state
def test_cloning_immutability(self):
clone = MancalaEnv.clone(self.game)
self.game.perform_move(Move(Side.SOUTH, 3))
self.assertEqual(clone.board.get_seeds(Side.SOUTH, 3), 7)
self.assertEqual(clone.side_to_move, Side.SOUTH)