add VDN fix

This commit is contained in:
romue 2021-06-22 14:19:13 +02:00
parent 543c2987e0
commit c5d677e9ba

View File

@ -39,9 +39,9 @@ class BaseBuffer:
return Experience(observations, next_observations, actions, rewards, dones)
class BaseDQN(nn.Module):
class BaseDDQN(nn.Module):
def __init__(self):
super(BaseDQN, self).__init__()
super(BaseDDQN, self).__init__()
self.net = nn.Sequential(
nn.Flatten(),
nn.Linear(3*5*5, 64),
@ -64,6 +64,27 @@ class BaseDQN(nn.Module):
return values + (advantages - advantages.mean())
class BaseDQN(nn.Module):
def __init__(self):
super(BaseDQN, self).__init__()
self.net = nn.Sequential(
nn.Flatten(),
nn.Linear(3*5*5, 64),
nn.ELU(),
nn.Linear(64, 64),
nn.ELU(),
nn.Linear(64, 9)
)
def act(self, x) -> np.ndarray:
with torch.no_grad():
action = self.forward(x).max(-1)[1].numpy()
return action
def forward(self, x):
return self.net(x)
def soft_update(local_model, target_model, tau):
# taken from https://github.com/BY571/Munchausen-RL/blob/master/M-DQN.ipynb
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
@ -154,10 +175,11 @@ class BaseQlearner:
print(f'Step: {step} ({(step/n_steps)*100:.2f}%)\tRunning reward: {sum(list(self.running_reward))/len(self.running_reward):.2f}\t'
f' eps: {self.eps:.4f}\tRunning loss: {sum(list(self.running_loss))/len(self.running_loss):.4f}\tUpdates:{self._n_updates}')
def _training_routine(self, obs, next_obs, action, reward):
def _training_routine(self, obs, next_obs, action):
current_q_values = self.q_net(obs)
current_q_values = torch.gather(current_q_values, dim=-1, index=action)
next_q_values_raw = self.target_q_net(next_obs).max(dim=-1)[0].reshape(-1, 1).detach()
#print(current_q_values.shape, next_q_values_raw.shape)
return current_q_values, next_q_values_raw
def _backprop_loss(self, loss):
@ -174,22 +196,21 @@ class BaseQlearner:
for _ in range(self.n_grad_steps):
experience = self.buffer.sample(self.batch_size, cer=self.train_every_n_steps)
if self.n_agents <= 1:
pred_q, target_q_raw = self._training_routine(experience.observation,
experience.next_observation,
experience.action,
experience.reward)
experience.next_observation,
experience.action)
else:
pred_q, target_q_raw, reward = [torch.zeros((self.batch_size, 1))]*3
pred_q, target_q_raw = torch.zeros((self.batch_size, 1)), torch.zeros((self.batch_size, 1))
for agent_i in range(self.n_agents):
q_values, next_q_values_raw = self._training_routine(experience.observation[:, agent_i],
experience.next_observation[:, agent_i],
experience.action[:, agent_i].unsqueeze(-1),
experience.reward)
experience.next_observation[:, agent_i],
experience.action[:, agent_i].unsqueeze(-1))
pred_q += q_values
target_q_raw += next_q_values_raw
target_q = experience.reward + (1 - experience.done) * self.gamma * target_q_raw
#print(pred_q[0], target_q_raw[0], target_q[0], experience.reward[0])
loss = torch.mean(self.reg_weight * pred_q + torch.pow(pred_q - target_q, 2))
self._backprop_loss(loss)
@ -245,7 +266,7 @@ class MDQN(BaseQlearner):
if __name__ == '__main__':
from environments.factory.simple_factory import SimpleFactory, DirtProperties, MovementProperties
N_AGENTS = 1
N_AGENTS = 2
dirt_props = DirtProperties(clean_amount=3, gain_amount=0.2, max_global_amount=30,
max_local_amount=5, spawn_frequency=1, max_spawn_ratio=0.05)
@ -255,6 +276,6 @@ if __name__ == '__main__':
env = SimpleFactory(dirt_properties=dirt_props, movement_properties=move_props, n_agents=N_AGENTS, pomdp_radius=2, max_steps=400, omit_agent_slice_in_obs=False, combin_agent_slices_in_obs=True)
dqn, target_dqn = BaseDQN(), BaseDQN()
learner = MDQN(dqn, target_dqn, env, BaseBuffer(40000), target_update=3500, lr=0.0007, gamma=0.99, n_agents=N_AGENTS, tau=0.95, max_grad_norm=10,
learner = BaseQlearner(dqn, target_dqn, env, BaseBuffer(40000), target_update=3500, lr=0.0007, gamma=0.99, n_agents=N_AGENTS, tau=0.95, max_grad_norm=10,
train_every_n_steps=4, eps_end=0.025, n_grad_steps=1, reg_weight=0.1, exploration_fraction=0.25, batch_size=64)
learner.learn(100000)