class NFSPAgent(object):
''' NFSP Agent implementation in TensorFlow.
'''
def __init__(self,
sess,
scope,
action_num=4,
state_shape=None,
hidden_layers_sizes=None,
reservoir_buffer_capacity=int(1e6),
anticipatory_param=0.1,
batch_size=256,
train_every=1,
rl_learning_rate=0.1,
sl_learning_rate=0.005,
min_buffer_size_to_learn=1000,
q_replay_memory_size=30000,
q_replay_memory_init_size=1000,
q_update_target_estimator_every=1000,
q_discount_factor=0.99,
q_epsilon_start=0.06,
q_epsilon_end=0,
q_epsilon_decay_steps=int(1e6),
q_batch_size=256,
q_train_every=1,
q_mlp_layers=None,
evaluate_with='average_policy'):
''' Initialize the NFSP agent.
Args:
sess (tf.Session): Tensorflow session object.
scope (string): The name scope of NFSPAgent.
action_num (int): The number of actions.
state_shape (list): The shape of the state space.
hidden_layers_sizes (list): The hidden layers sizes for the layers of
the average policy.
reservoir_buffer_capacity (int): The size of the buffer for average policy.
anticipatory_param (float): The hyper-parameter that balances rl/avarage policy.
batch_size (int): The batch_size for training average policy.
train_every (int): Train the SL policy every X steps.
rl_learning_rate (float): The learning rate of the RL agent.
sl_learning_rate (float): the learning rate of the average policy.
min_buffer_size_to_learn (int): The minimum buffer size to learn for average policy.
q_replay_memory_size (int): The memory size of inner DQN agent.
q_replay_memory_init_size (int): The initial memory size of inner DQN agent.
q_update_target_estimator_every (int): The frequency of updating target network for
inner DQN agent.
q_discount_factor (float): The discount factor of inner DQN agent.
q_epsilon_start (float): The starting epsilon of inner DQN agent.
q_epsilon_end (float): the end epsilon of inner DQN agent.
q_epsilon_decay_steps (int): The decay steps of inner DQN agent.
q_batch_size (int): The batch size of inner DQN agent.
q_train_step (int): Train the model every X steps.
q_mlp_layers (list): The layer sizes of inner DQN agent.
evaluate_with (string): The value can be 'best_response' or 'average_policy'
'''
self.use_raw = False
self._sess = sess
self._scope = scope
self._action_num = action_num
self._state_shape = state_shape
self._layer_sizes = hidden_layers_sizes
self._batch_size = batch_size
self._train_every = train_every
self._sl_learning_rate = sl_learning_rate
self._anticipatory_param = anticipatory_param
self._min_buffer_size_to_learn = min_buffer_size_to_learn
self._reservoir_buffer = ReservoirBuffer(reservoir_buffer_capacity)
self._prev_timestep = None
self._prev_action = None
self.evaluate_with = evaluate_with
# Total timesteps
self.total_t = 0
# Step counter to keep track of learning.
self._step_counter = 0
with tf.compat.v1.variable_scope(scope):
# Inner RL agent
self._rl_agent = DQNAgent(sess, scope+'_dqn', q_replay_memory_size, q_replay_memory_init_size, q_update_target_estimator_every, q_discount_factor, q_epsilon_start, q_epsilon_end, q_epsilon_decay_steps, q_batch_size, action_num, state_shape, q_train_every, q_mlp_layers, rl_learning_rate)
with tf.compat.v1.variable_scope('sl'):
# Build supervised model
self._build_model()
self.sample_episode_policy()
def _build_model(self):
''' build the model for supervised learning
'''
# Placeholders.
input_shape = [None]
input_shape.extend(self._state_shape)
self._info_state_ph = tf.compat.v1.placeholder(
shape=input_shape,
dtype=tf.float32)
self._X = tf.contrib.layers.flatten(self._info_state_ph)
# Boolean to indicate whether is training or not
self.is_train = tf.compat.v1.placeholder(tf.bool, name="is_train");
# Batch Normalization
self._X = tf.layers.batch_normalization(self._X, training=True)
self._action_probs_ph = tf.compat.v1.placeholder(
shape=[None, self._action_num], dtype=tf.float32)
# Average policy network.
fc = self._X
for dim in self._layer_sizes:
fc = tf.contrib.layers.fully_connected(fc, dim, activation_fn=tf.tanh)
self._avg_policy = tf.contrib.layers.fully_connected(fc, self._action_num, activation_fn=None)
self._avg_policy_probs = tf.nn.softmax(self._avg_policy)
# Loss
self._loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(
labels=tf.stop_gradient(self._action_probs_ph),
logits=self._avg_policy))
optimizer = tf.train.AdamOptimizer(learning_rate=self._sl_learning_rate, name='nfsp_adam')
update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS, scope=tf.compat.v1.get_variable_scope().name)
with tf.control_dependencies(update_ops):
self._learn_step = optimizer.minimize(self._loss)
def feed(self, ts):
''' Feed data to inner RL agent
Args:
ts (list): A list of 5 elements that represent the transition.
'''
self._rl_agent.feed(ts)
self.total_t += 1
if self.total_t>0 and len(self._reservoir_buffer) >= self._min_buffer_size_to_learn and self.total_t%self._train_every == 0:
sl_loss = self.train_sl()
print('\rINFO - Agent {}, step {}, sl-loss: {}'.format(self._scope, self.total_t, sl_loss), end='')
def step(self, state):
''' Returns the action to be taken.
Args:
state (dict): The current state
Returns:
action (int): An action id
'''
obs = state['obs']
legal_actions = state['legal_actions']
if self._mode == MODE.best_response:
probs = self._rl_agent.predict(obs)
one_hot = np.eye(len(probs))[np.argmax(probs)]
self._add_transition(obs, one_hot)
elif self._mode == MODE.average_policy:
probs = self._act(obs)
probs = remove_illegal(probs, legal_actions)
action = np.random.choice(len(probs), p=probs)
return action
def eval_step(self, state):
''' Use the average policy for evaluation purpose
Args:
state (dict): The current state.
Returns:
action (int): An action id.
probs (list): The list of action probabilies
'''
if self.evaluate_with == 'best_response':
action, probs = self._rl_agent.eval_step(state)
elif self.evaluate_with == 'average_policy':
obs = state['obs']
legal_actions = state['legal_actions']
probs = self._act(obs)
probs = remove_illegal(probs, legal_actions)
action = np.random.choice(len(probs), p=probs)
else:
raise ValueError("'evaluate_with' should be either 'average_policy' or 'best_response'.")
return action, probs
def sample_episode_policy(self):
''' Sample average/best_response policy
'''
if np.random.rand() < self._anticipatory_param:
self._mode = MODE.best_response
else:
self._mode = MODE.average_policy
def _act(self, info_state):
''' Predict action probability givin the observation and legal actions
Args:
info_state (numpy.array): An obervation.
Returns:
action_probs (numpy.array): The predicted action probability.
'''
info_state = np.expand_dims(info_state, axis=0)
action_probs = self._sess.run(
self._avg_policy_probs,
feed_dict={self._info_state_ph: info_state, self.is_train: False})[0]
return action_probs
def _add_transition(self, state, probs):
''' Adds the new transition to the reservoir buffer.
Transitions are in the form (state, probs).
Args:
state (numpy.array): The state.
probs (numpy.array): The probabilities of each action.
'''
transition = Transition(
info_state=state,
action_probs=probs)
self._reservoir_buffer.add(transition)
def train_sl(self):
''' Compute the loss on sampled transitions and perform a avg-network update.
If there are not enough elements in the buffer, no loss is computed and
`None` is returned instead.
Returns:
loss (float): The average loss obtained on this batch of transitions or `None`.
'''
if (len(self._reservoir_buffer) < self._batch_size or
len(self._reservoir_buffer) < self._min_buffer_size_to_learn):
return None
transitions = self._reservoir_buffer.sample(self._batch_size)
info_states = [t.info_state for t in transitions]
action_probs = [t.action_probs for t in transitions]
loss, _ = self._sess.run(
[self._loss, self._learn_step],
feed_dict={
self._info_state_ph: info_states,
self._action_probs_ph: action_probs,
self.is_train: True,
})
return loss
sess.run(tf.global_variables_initializer())
# Init a Logger to plot the learning curve
for episode in range(episode_num):
print(episode)
sys.stdout.flush()
# Generate data from the environment
trajectories, _ = train_env.run(is_training=True)
# Feed transitions into agent memory, and train the agent
for tss in trajectories[:1]:
for ts in tss:
agent.feed(ts)
#print(episode)
if episode % evaluate_every == 0:
eval_env = rlcard.make('doudizhu', config={'seed': 0, 'allow_step_back': True})
eval_env.set_agents([agent, SRandomAgent(eval_env.action_num, seed=0), SRandomAgent(eval_env.action_num, seed=0)])
time_start = time.time()
payoffs1 = general_tournament(eval_env,evaluate_num,False)
logger.log("episode:{} time:{} landlord winrate:{}".format(episode,time.time()-time_start,payoffs1[0]))
L_WR_logger.log_performance(episode,payoffs1[0])
eval_env = rlcard.make('doudizhu', config={'seed': 0, 'allow_step_back': True})
eval_env.set_agents([SRandomAgent(eval_env.action_num, seed=0), SRandomAgent(eval_env.action_num, seed=0),agent])
time_start = time.time()
# Initialize global variables
sess.run(tf.global_variables_initializer())
# Init a Logger to plot the learning curve
for episode in range(episode_num):
# Generate data from the environment
trajectories, _ = train_env.run(is_training=True)
# Feed transitions into agent memory, and train the agent
loss = 0.0
loss_count = 0
for tss in trajectories:
for ts in tss:
cur_loss = agent.feed(ts)
if cur_loss > 0:
loss += cur_loss
loss_count += 1
if loss_count > 0:
loss = loss / loss_count
#print(episode)
if episode % evaluate_every == 0:
eval_env = rlcard.make('doudizhu',
config={
'seed': 0,
'allow_step_back': True
})
eval_env.set_agents([
agent,
