def evaluate_players(p1: Player,
p2: Player,
games_per_battle=100,
num_battles=100):
board = Board()
p1_wins = []
p2_wins = []
draws = []
game_number = []
game_counter = 0
TFSessionManager.set_session(tf.Session())
TFSessionManager.get_session().run(tf.global_variables_initializer())
for i in range(num_battles):
p1win, p2win, draw = battle(p1, p2, games_per_battle, False)
p1_wins.append(p1win)
p2_wins.append(p2win)
draws.append(draw)
game_counter = game_counter + 1
game_number.append(game_counter)
TFSessionManager.set_session(None)
return game_number, p1_wins, p2_wins, draws
def final_result(self, result: GameResult):
"""
This method is called once the game is over. If `self.training` is True, we execute a training run for
the Neural Network.
:param result: The result of the game that just finished.
"""
# Compute the final reward based on the game outcome
if (result == GameResult.NAUGHT_WIN
and self.side == NAUGHT) or (result == GameResult.CROSS_WIN
and self.side == CROSS):
reward = self.win_value # type: float
elif (result == GameResult.NAUGHT_WIN
and self.side == CROSS) or (result == GameResult.CROSS_WIN
and self.side == NAUGHT):
reward = self.loss_value # type: float
elif result == GameResult.DRAW:
reward = self.draw_value # type: float
else:
raise ValueError("Unexpected game result {}".format(result))
# The final reward is also the Q value we want to learn for the action that led to it.
self.next_max_log.append(reward)
# If we are in training mode we run the optimizer.
if self.training:
# We calculate our new estimate of what the true Q values are and feed that into the network as
# learning target
targets = self.calculate_targets()
# We convert the input states we have recorded to feature vectors to feed into the training.
nn_input = [
self.board_state_to_nn_input(x)
for x in self.board_position_log
]
# We run the training step with the recorded inputs and new Q value targets.
TFSN.get_session().run([self.nn.train_step],
feed_dict={
self.nn.input_positions: nn_input,
self.nn.target_input: targets
})
self.random_move_prob *= self.random_move_decrease
def get_probs(self, input_pos: [np.ndarray]) -> ([float], [float]):
"""
Compute action probabilities through the Neural Network
:param input_pos: List of input states for which to compute probabilities
:return: Tuple of lists of action probabilities and raw logits for the given input states
"""
probs, logits = TFSN.get_session().run(
[self.nn.output, self.nn.logits],
feed_dict={self.nn.state_in: input_pos})
return probs, logits
def get_probs(self, input_pos: np.ndarray) -> ([float], [float]):
"""
Feeds the feature vector `input_pos` which encodes a board state into the Neural Network and computes the
Q values and corresponding probabilities for all moves (including illegal ones).
:param input_pos: The feature vector to be fed into the Neural Network.
:return: A tuple of probabilities and q values of all actions (including illegal ones).
"""
probs, qvalues = TFSN.get_session().run(
[self.nn.probabilities, self.nn.q_values],
feed_dict={self.nn.input_positions: [input_pos]})
return probs[0], qvalues[0]
def get_probs(self, input_pos: [np.ndarray],
network: QNetwork) -> ([float], [float]):
"""
Feeds the feature vectors `input_pos` (which encode a board states) into the Neural Network and computes the
Q values and corresponding probabilities for all moves (including illegal ones).
:param network: The network to get probabilities from
:param input_pos: A list of feature vectors to be fed into the Neural Network.
:return: A list of tuples of probabilities and q values of all actions (including illegal ones).
"""
probs, qvalues = TFSN.get_session().run(
[network.probabilities, network.q_values],
feed_dict={network.input_positions: input_pos})
return probs, qvalues
from tic_tac_toe.DirectPolicyAgent import DirectPolicyAgent
min_reward = -3
max_reward = 3
num_reward_steps = 1 + max_reward - min_reward
rewards = np.zeros((num_reward_steps, num_reward_steps))
for loss_reward in range(min_reward, max_reward):
for draw_reward in range(loss_reward + 1, max_reward + 1):
tf.reset_default_graph()
TFSessionManager.set_session(tf.Session())
sess = TFSessionManager.get_session()
nnplayer = DirectPolicyAgent("PolicyLearner1",
loss_value=loss_reward,
draw_value=draw_reward)
rm_player = RndMinMaxAgent()
sess.run(tf.global_variables_initializer())
game_number, p1_wins, p2_wins, draws = evaluate_players(
nnplayer, rm_player, num_battles=1000,
silent=True) # , num_battles = 20)
print("With loss reward {} and draw reward {} we get draws: {}".format(
loss_reward, draw_reward, draws[-1]))
def final_result(self, result: GameResult):
"""
This method is called once the game is over. If `self.training` is True, we execute a training run for
the Neural Network.
:param result: The result of the game that just finished.
"""
self.game_counter += 1
# Compute the final reward based on the game outcome
if (result == GameResult.NAUGHT_WIN
and self.side == NAUGHT) or (result == GameResult.CROSS_WIN
and self.side == CROSS):
reward = self.win_value # type: float
elif (result == GameResult.NAUGHT_WIN
and self.side == CROSS) or (result == GameResult.CROSS_WIN
and self.side == NAUGHT):
reward = self.loss_value # type: float
elif result == GameResult.DRAW:
reward = self.draw_value # type: float
else:
raise ValueError("Unexpected game result {}".format(result))
self.add_game_to_replay_buffer(reward)
# If we are in training mode we run the optimizer.
if self.training and (self.game_counter > self.pre_training_games):
batch_third = self.batch_size // 3
train_batch = self.replay_buffer_win.sample(batch_third)
train_batch.extend(self.replay_buffer_loss.sample(batch_third))
train_batch.extend(self.replay_buffer_draw.sample(batch_third))
train_batch = np.array(train_batch)
#
# Let's compute the target q values for all non terminal move
# We extract the resulting state, run it through the target net work and
# get the maximum q value (of all valid moves)
next_states = [s[2] for s in train_batch if s[2] is not None]
target_qs = []
if len(next_states) > 0:
probs, qvals = self.get_valid_probs(
[self.board_state_to_nn_input(s) for s in next_states],
self.target_net, [Board(s) for s in next_states])
i = 0
for t in train_batch:
if t[2] is not None:
max_move = np.argmax(probs[i])
max_qval = qvals[i][max_move]
target_qs.append(max_qval * self.reward_discount)
i += 1
else:
target_qs.append(t[3])
if i != len(next_states):
print("Something wrong here!!!")
else:
target_qs.extend(train_batch[:, 3])
# We convert the input states we have recorded to feature vectors to feed into the training.
nn_input = [
self.board_state_to_nn_input(x[0]) for x in train_batch
]
actions = train_batch[:, 1]
# We run the training step with the recorded inputs and new Q value targets.
summary, _ = TFSN.get_session().run(
[self.q_net.merge, self.q_net.train_step],
feed_dict={
self.q_net.input_positions: nn_input,
self.q_net.target_q: target_qs,
self.q_net.actions: actions
})
self.random_move_prob *= self.random_move_decrease
if self.writer is not None:
self.writer.add_summary(summary, self.game_counter)
summary = tf.Summary(value=[
tf.Summary.Value(tag='Random_Move_Probability',
simple_value=self.random_move_prob)
])
self.writer.add_summary(summary, self.game_counter)
TFSN.get_session().run(self.graph_copy_op)
def final_result(self, result: GameResult):
"""
Called when the game has ended. Time to record results and train the network.
:param result: The final result of the game
"""
# Compute the final reward based on the game outcome
if (result == GameResult.NAUGHT_WIN
and self.side == NAUGHT) or (result == GameResult.CROSS_WIN
and self.side == CROSS):
final_reward = self.win_value # type: float
elif (result == GameResult.NAUGHT_WIN
and self.side == CROSS) or (result == GameResult.CROSS_WIN
and self.side == NAUGHT):
final_reward = self.loss_value # type: float
elif result == GameResult.DRAW:
final_reward = self.draw_value # type: float
else:
raise ValueError("Unexpected game result {}".format(result))
self.game_counter += 1
rewards = self.calculate_rewards(final_reward, len(self.action_log))
# noinspection PyTypeChecker
self.add_game_to_replay_buffer(final_reward, rewards)
# If we are in training mode we run the optimizer.
if self.training and (self.game_counter > self.pre_training_games):
batch_third = self.batch_size // 3
train_batch = self.replay_buffer_win.sample(batch_third)
train_batch.extend(self.replay_buffer_loss.sample(batch_third))
train_batch.extend(self.replay_buffer_draw.sample(batch_third))
train_batch = np.array(train_batch)
# We convert the input states we have recorded to feature vectors to feed into the training.
nn_input = np.array(
[self.board_state_to_nn_input(x[0]) for x in train_batch])
actions = np.array(train_batch[:, 1])
rewards = np.array(train_batch[:, 2])
feed_dict = {
self.nn.reward_holder: rewards,
self.nn.action_holder: actions,
self.nn.state_in: nn_input
}
summary, _, inds, rps, loss = TFSN.get_session().run(
[
self.nn.merge, self.nn.update_batch, self.nn.indexes,
self.nn.responsible_outputs, self.nn.loss
],
feed_dict=feed_dict)
self.random_move_probability *= self.random_move_decrease
if self.writer is not None:
self.writer.add_summary(summary, self.game_counter)
summary = tf.Summary(value=[
tf.Summary.Value(tag='Random_Move_Probability',
simple_value=self.random_move_probability)
])
self.writer.add_summary(summary, self.game_counter)
player1 = RandomPlayer()
player2 = RandomPlayer()
p1_wins = []
p1count = 0
p2_wins = []
p2count = 0
draws = []
drawcount = 0
count = []
num_battles = 100
games_per_battle = 10
TFSessionManager.set_session(tf.Session())
TFSessionManager.get_session().run(tf.global_variables_initializer())
for i in range(num_battles):
p1win, p2win, draw = battle(player1, player2, games_per_battle, True)
p1_wins.append(p1win)
p1count += p1win
p2_wins.append(p2win)
p2count += p2win
draws.append(draw)
drawcount += draw
count.append(i * games_per_battle)
#TFSessionManager.set_session(None)
p = plt.plot(count, draws, 'r-', count, p1_wins, 'g-', count, p2_wins, 'b-')
print("p1 wins = " + str(p1count / (num_battles * games_per_battle) * 100) +
"% (green)")
print("p2 wins = " + str(p2count / (num_battles * games_per_battle) * 100) +
mm_player = MinMaxAgent()
tq_player = TQPlayer()
p1_wins = []
p2_wins = []
draws = []
game_number = []
game_counter = 0
num_battles = 10
games_per_battle = 100
num_training_battles = 1000
TFSessionManager.set_session(tf.Session())
TFSessionManager.get_session().run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('log', TFSessionManager.get_session().graph)
# nnplayer rndplayer mm_player
p1_t = deep_nnplayer
p2_t = mm_player
p1 = p1_t
p2 = p2_t
# nnplayer.training= False
# nnplayer2.training= False
for i in range(num_training_battles):
p1win, p2win, draw = battle(p1_t, p2_t, games_per_battle, False)
p1_wins.append(p1win)