done = 1
if done:
# Record total reward
EPs_total_reward.append(EP_reward_sum)
states = np.vstack(states)
next_states = np.vstack(next_states)
actions = np.vstack(actions)
values = get_return(values,0.95)
values = np.vstack(values)
for ve in range(CRITIC_EPOCHS):
critic_loss = tuc.train_critic(states, values, actions)
states = []
next_states = []
actions = []
values = []
# Print total reward
print("PG episode : {0: <5} , total reward : {1: <5}".format(EP,EP_reward_sum))
# Stop to train agent
PG_agent.agent_REINFORCE()
break
# Move to the next state
state = next_state
# Perform the optimization
if done:
states = torch.cat(agent.states)
values = torch.tensor(np.expand_dims(np.array(
agent.get_values()),
axis=1),
dtype=torch.float32)
actions_matrix = torch.cat(actions_matrix, 0).cuda()
#print(len(actions_matrix))
for e in range(3):
TUC_dynamic.train_critic(states.cuda(), values.cuda(),
actions_matrix.cuda())
print("updating agent !")
agent.REINFORCE()
break
#==================== loop of training procedure ==========================================#
time_cost = time.time() - now
print('epoch = %d, time_cost = %.4f' % (epoch, time_cost))
# save the whole model
agent.save_model("./model_final/pg_TUC_agent_2")
TUC_dynamic.save_model("./model_final/pg_TUC_2")
print('Complete')