for i in range(num_runs):
vel = np.zeros(num_steps)
# logging.info("Iter #" + str(i))
print('episode is:', i)
ret = 0
ret_list = []
state = env.reset()
aset = []
for j in range(num_steps):
# manager actions
# one controllers
a = agent.choose_action([state])
aset.append(a)
next_state, reward, done, _ = env.step(a)
# if j%300==0:
# print(state)
# print(a)
ret += reward
ret_list.append(reward)
if agent.num_experience > 2000:
ploss, qloss, reg_loss, Q = agent.learn(batch_size=32)
if done:
agent.remember(state, a, reward, next_state, 1.)
break
agent.remember(state, a, reward, next_state, 0.)
state = next_state[:]
action_dict = {}
a = sess.run(agent.action,
feed_dict={
agent.adj: [adj],
agent.state_holder: state_,
agent.vecholder: np.asarray([vec])
})
k = 0
for key, value in state.items():
action_dict[key] = a[k]
k += 1
aset.append(a)
next_state, reward, done, _ = env.step(action_dict)
next_state_ = np.array(list(next_state.values())).reshape(
1, -1).tolist()
rewards = list(reward.values())
## calculate individual reward
# for k in range(len(rewards)):
# print('agent',k,'reward is:',rewards[k])
ret += np.average(
rewards) ##here we consider the rewards of each agent
ret_list.append(rewards)
agent.remember(state_, a, rewards, next_state_, 0., adj)
if agent.num_experience > 200: ##could change to 200
ploss, qloss = agent.learn(batch_size=32)
