"""
DCG: Deep coordination graphs
Paper link: http://proceedings.mlr.press/v119/boehmer20a/boehmer20a.pdf
Implementation: TensorFlow 2.X
"""
from argparse import Namespace
from xuance.tensorflow import tf, tk, Module
from xuance.tensorflow.learners import LearnerMAS
[docs]
class DCG_Learner(LearnerMAS):
def __init__(self,
config: Namespace,
policy: Module,
optimizer: tk.optimizers.Optimizer,
device: str = "cpu:0",
model_dir: str = "./",
gamma: float = 0.99,
sync_frequency: int = 100
):
self.gamma = gamma
self.use_rnn = config.use_rnn
self.sync_frequency = sync_frequency
self.dim_hidden_state = policy.representation.output_shapes['state'][0]
self.sync_frequency = sync_frequency
super(DCG_Learner, self).__init__(config, policy, optimizer, device, model_dir)
self.mse_loss = tk.losses.MeanSquaredError()
[docs]
def get_hidden_states(self, obs_n, *rnn_hidden, use_target_net=False):
if self.use_rnn:
if use_target_net:
outputs = self.policy.target_representation(obs_n, *rnn_hidden)
else:
outputs = self.policy.representation(obs_n, *rnn_hidden)
hidden_states = outputs['state']
rnn_hidden = (outputs['rnn_hidden'], outputs['rnn_cell'])
else:
shape_obs_n = obs_n.shape
rep_in = tf.reshape(obs_n, [-1, shape_obs_n[-1]])
if use_target_net:
hidden_states = self.policy.target_representation(rep_in)['state']
else:
hidden_states = self.policy.representation(rep_in)['state']
hidden_states_out = tf.reshape(hidden_states, shape_obs_n[:-1] + (self.dim_hidden_state, ))
rnn_hidden = None
return rnn_hidden, hidden_states_out
[docs]
def get_graph_values(self, hidden_states, use_target_net=False):
if use_target_net:
utilities = self.policy.target_utility(hidden_states)
payoff = self.policy.target_payoffs(hidden_states, self.policy.graph.edges_from, self.policy.graph.edges_to)
else:
utilities = self.policy.utility(hidden_states)
payoff = self.policy.payoffs(hidden_states, self.policy.graph.edges_from.numpy(), self.policy.graph.edges_to.numpy())
return utilities, payoff
[docs]
def act(self, hidden_states, avail_actions=None):
with torch.no_grad():
f_i, f_ij = self.get_graph_values(hidden_states)
n_edges = self.policy.graph.n_edges
n_vertexes = self.policy.graph.n_vertexes
f_i_mean = tf.cast(f_i, dtype=tf.double) / n_vertexes
f_ij_mean = tf.cast(f_ij, dtype=tf.double) / n_edges
f_ji_mean = copy.deepcopy(tf.transpose(f_ij_mean, perm=(0, 1, 3, 2)))
batch_size = f_i.shape[0]
msg_ij = torch.zeros(batch_size, n_edges, self.dim_act) # i -> j (send)
msg_ji = torch.zeros(batch_size, n_edges, self.dim_act) # j -> i (receive)
#
msg_forward = torch_scatter.scatter_add(src=msg_ij, index=self.policy.graph.edges_to, dim=1,
dim_size=n_vertexes)
msg_backward = torch_scatter.scatter_add(src=msg_ji, index=self.policy.graph.edges_from, dim=1,
dim_size=n_vertexes)
f_i_mean = torch.tensor(f_i_mean.numpy())
f_ij_mean = torch.tensor(f_ij_mean.numpy())
f_ji_mean = torch.tensor(f_ji_mean.numpy())
utility = f_i_mean + msg_forward + msg_backward
if len(self.policy.graph.edges) != 0:
for i in range(self.args.n_msg_iterations):
joint_forward = (utility[:, self.policy.graph.edges_from, :] - msg_ji).unsqueeze(dim=-1) + f_ij_mean
joint_backward = (utility[:, self.policy.graph.edges_to, :] - msg_ij).unsqueeze(dim=-1) + f_ji_mean
msg_ij = joint_forward.max(dim=-2).values
msg_ji = joint_backward.max(dim=-2).values
if self.args.msg_normalized:
msg_ij -= msg_ij.mean(dim=-1, keepdim=True)
msg_ji -= msg_ji.mean(dim=-1, keepdim=True)
msg_forward = torch_scatter.scatter_add(src=msg_ij, index=self.policy.graph.edges_to, dim=1,
dim_size=n_vertexes)
msg_backward = torch_scatter.scatter_add(src=msg_ji, index=self.policy.graph.edges_from, dim=1,
dim_size=n_vertexes)
utility = f_i_mean + msg_forward + msg_backward
if avail_actions is not None:
avail_actions = torch.Tensor(avail_actions)
utility_detach = utility.clone().detach()
utility_detach[avail_actions == 0] = -1e10
actions_greedy = utility_detach.argmax(dim=-1)
else:
actions_greedy = utility.argmax(dim=-1)
return actions_greedy
[docs]
def q_dcg(self, hidden_states, actions, states=None, use_target_net=False):
f_i, f_ij = self.get_graph_values(hidden_states, use_target_net=use_target_net)
f_i_mean = tf.cast(f_i, tf.double) / self.policy.graph.n_vertexes
f_ij_mean = tf.cast(f_ij, tf.double) / self.policy.graph.n_edges
utilities = tf.reduce_sum(tf.gather(f_i_mean, tf.expand_dims(actions, -1), axis=-1, batch_dims=-1), axis=1)
if len(self.policy.graph.edges) == 0 or self.args.n_msg_iterations == 0:
return utilities
edges_from = self.policy.graph.edges_from.numpy()
edges_to = self.policy.graph.edges_to.numpy()
actions_ij = tf.expand_dims(tf.gather(actions, edges_from, axis=1) * self.dim_act + tf.gather(actions, edges_to, axis=1), -1)
payoffs = tf.reduce_sum(tf.gather(tf.reshape(f_ij_mean, list(f_ij_mean.shape[0:-2]) + [-1]), actions_ij, axis=-1, batch_dims=-1), axis=1)
if self.args.agent == "DCG_S":
state_value = self.policy.bias(states)
return utilities + payoffs + state_value
else:
return utilities + payoffs
[docs]
def update(self, sample):
self.iterations += 1
with tf.device(self.device):
state = tf.convert_to_tensor(sample['state'])
state_next = tf.convert_to_tensor(sample['state_next'])
obs = tf.convert_to_tensor(sample['obs'])
actions = tf.convert_to_tensor(sample['actions'], dtype=tf.int64)
obs_next = tf.convert_to_tensor(sample['obs_next'])
rewards = tf.reduce_mean(tf.convert_to_tensor(sample['rewards']), axis=1)
terminals = tf.reshape(tf.convert_to_tensor(sample['terminals'].all(axis=-1, keepdims=True), dtype=tf.float32), [-1, 1])
agent_mask = tf.reshape(tf.convert_to_tensor(sample['agent_mask'], dtype=tf.float32),
[-1, self.n_agents, 1])
IDs = tf.tile(tf.expand_dims(tf.eye(self.n_agents), axis=0), multiples=(self.args.batch_size, 1, 1))
batch_size = obs.shape[0]
with tf.GradientTape() as tape:
_, hidden_states = self.get_hidden_states(obs, use_target_net=False)
q_eval_a = self.q_dcg(hidden_states, actions, states=state, use_target_net=False)
_, hidden_states_next = self.get_hidden_states(obs_next)
action_next_greedy = tf.convert_to_tensor(self.act(hidden_states_next))
_, hidden_states_target = self.get_hidden_states(obs_next, use_target_net=True)
q_next_a = self.q_dcg(hidden_states_target, action_next_greedy, states=state_next, use_target_net=True)
q_next_a = tf.cast(q_next_a, dtype=tf.float32)
q_target = rewards + (1 - terminals) * self.args.gamma * q_next_a
# calculate the loss function
y_true = tf.stop_gradient(tf.reshape(q_target, [-1]))
y_pred = tf.reshape(q_eval_a, [-1])
loss = self.mse_loss(y_true, y_pred)
gradients = tape.gradient(loss, self.policy.trainable_variables)
self.optimizer.apply_gradients([
(grad, var)
for (grad, var) in zip(gradients, self.policy.trainable_variables)
if grad is not None
])
if self.iterations % self.sync_frequency == 0:
self.policy.copy_target()
lr = self.optimizer._decayed_lr(tf.float32)
info = {
"learning_rate": lr.numpy(),
"loss_Q": loss.numpy(),
"predictQ": tf.math.reduce_mean(q_eval_a).numpy()
}
return info