def test_train(self):
memory_init_size = 100
step_num = 1500
agent = DQNAgent(scope='dqn',
replay_memory_size=500,
replay_memory_init_size=memory_init_size,
update_target_estimator_every=100,
state_shape=[2],
mlp_layers=[10, 10],
device=torch.device('cpu'))
predicted_action, _ = agent.eval_step({
'obs':
np.random.random_sample((2, )),
'legal_actions': [0, 1]
})
self.assertGreaterEqual(predicted_action, 0)
self.assertLessEqual(predicted_action, 1)
for _ in range(step_num):
ts = [{
'obs': np.random.random_sample((2, )),
'legal_actions': [0, 1]
},
np.random.randint(2), 0, {
'obs': np.random.random_sample((2, )),
'legal_actions': [0, 1]
}, True]
agent.feed(ts)
state_dict = agent.get_state_dict()
self.assertIsInstance(state_dict, dict)
predicted_action = agent.step({
'obs': np.random.random_sample((2, )),
'legal_actions': [0, 1]
})
self.assertGreaterEqual(predicted_action, 0)
self.assertLessEqual(predicted_action, 1)
class NFSPAgent(object):
''' An approximate clone of rlcard3.agents.nfsp_agent that uses
pytorch instead of tensorflow. Note that this implementation
differs from Henrich and Silver (2016) in that the supervised
training minimizes cross-entropy with respect to the stored
action probabilities rather than the realized actions.
'''
def __init__(self,
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',
device=None):
''' Initialize the NFSP agent.
Args:
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.
device (torch.device): Whether to use the cpu or gpu
'''
self.use_raw = False
self._scope = scope
self._action_num = action_num
self._state_shape = state_shape
self._layer_sizes = hidden_layers_sizes + [action_num]
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
if device is None:
self.device = torch.device(
'cuda:0' if torch.cuda.is_available() else 'cpu')
else:
self.device = device
# Total timesteps
self.total_t = 0
# Step counter to keep track of learning.
self._step_counter = 0
# Build the action-value network
self._rl_agent = DQNAgent(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, device)
# Build the average policy supervised model
self._build_model()
self.sample_episode_policy()
def _build_model(self):
''' Build the average policy network
'''
# configure the average policy network
policy_network = AveragePolicyNetwork(self._action_num,
self._state_shape,
self._layer_sizes)
policy_network = policy_network.to(self.device)
self.policy_network = policy_network
self.policy_network.eval()
# xavier init
for p in self.policy_network.parameters():
if len(p.data.shape) > 1:
nn.init.xavier_uniform_(p.data)
# configure optimizer
self.policy_network_optimizer = torch.optim.Adam(
self.policy_network.parameters(), lr=self._sl_learning_rate)
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)
self._add_transition(obs, probs)
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.
'''
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
Not connected to computation graph
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)
info_state = torch.from_numpy(info_state).float().to(self.device)
with torch.no_grad():
log_action_probs = self.policy_network(info_state).cpu().numpy()
action_probs = np.exp(log_action_probs)[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]
self.policy_network_optimizer.zero_grad()
self.policy_network.train()
# (batch, state_size)
info_states = torch.from_numpy(np.array(info_states)).float().to(
self.device)
# (batch, action_num)
eval_action_probs = torch.from_numpy(
np.array(action_probs)).float().to(self.device)
# (batch, action_num)
log_forecast_action_probs = self.policy_network(info_states)
ce_loss = -(eval_action_probs *
log_forecast_action_probs).sum(dim=-1).mean()
ce_loss.backward()
self.policy_network_optimizer.step()
ce_loss = ce_loss.item()
self.policy_network.eval()
return ce_loss
def get_state_dict(self):
''' Get the state dict to save models
Returns:
(dict): A dict of model states
'''
state_dict = self._rl_agent.get_state_dict()
state_dict[self._scope] = self.policy_network.state_dict()
return state_dict
def load(self, checkpoint):
''' Load model
Args:
checkpoint (dict): the loaded state
'''
self.policy_network.load_state_dict(checkpoint[self._scope])