def AC_main():
#agent = PGagent(agentParam)
all_rw = []
#n_agents = 1#2
if mode == "social":
multiPG = socialAgents([
PGagent(env_dim["cleanup"][0], env_dim["cleanup"][1], add_para(i))
for i in range(n_agents)
], agentParam)
elif mode == "AC":
multiPG = AC_Agents([IAC(8, 400) for i in range(n_agents)])
else:
multiPG = independentAgent([
PGagent(env_dim["cleanup"][0], env_dim["cleanup"][1], add_para(i))
for i in range(n_agents)
])
for i_episode in range(n_episode):
n_state, ep_reward = env.reset(), 0
for t in range(n_steps):
if mode == "social":
actions = multiPG.select_mask_actions(n_state)
else:
actions = multiPG.select_actions(
n_state) ##agent.select_action(state)
#actions = [ random.randint(0,7) for i in range(n_agents)]
n_state_, n_reward, _, _ = env.step(actions)
if render and i_episode == 1:
env.render(impath, t)
#multiPG.push_reward(n_reward)
ep_reward += sum(n_reward)
multiPG.update(n_state, n_reward, n_state_, actions)
running_reward = ep_reward
#if test_mode == False:
# multiPG.update_agents()
all_rw.append(ep_reward)
if i_episode % (args.log_interval * 2) == 0 and ifsave_data:
np.save("data/" + model_name + ".npy", all_rw)
if i_episode % args.log_interval == 0:
print('Episode {}\tLast reward: {:.2f}\tAverage reward: {:.2f}'.
format(i_episode, ep_reward, running_reward))
if i_episode % save_eps == 0 and i_episode > 10 and ifsave_model:
multiPG.save(file_name)
def AC_main():
#agent = PGagent(agentParam)
all_rw = []
#n_agents = 1#2
if mode == "social":
multiPG = socialAgents([
PGagent(env_dim["cleanup"][0], env_dim["cleanup"][1], add_para(i))
for i in range(n_agents)
], agentParam)
elif mode == "AC":
multiPG = Agents([Centralised_AC(4, 100) for i in range(n_agents)], 50)
else:
multiPG = independentAgent([
PGagent(env_dim["cleanup"][0], env_dim["cleanup"][1], add_para(i))
for i in range(n_agents)
])
for i_episode in range(1000):
n_state, ep_reward = env.reset(), 0
n_state = n_state[0]
test_reward_sum = 0
for t in range(1000):
if mode == "social":
actions = multiPG.select_mask_actions(n_state)
else:
actions = multiPG.choose_action(
process_state(n_state)) ##agent.select_action(state)
#actions = [ random.randint(0,7) for i in range(n_agents)]
a = deepcopy(actions)
for i in range(len(actions)):
a[i] = [actions[i][0]]
n_state_, n_reward, _, _, test_reward = env.step(a)
test_reward_sum += test_reward
if render and i_episode != 1:
# env.render(impath,t)
env.render()
# time.sleep(0.05)
#multiPG.push_reward(n_reward)
ep_reward += sum(n_reward)
# if [1] in process_state(n_reward):
# print("i_episode %d:"%i_episode,process_state(n_reward))
multiPG.update(process_state(n_state), process_state(n_reward),
process_state(n_state_), actions)
n_state = n_state_
running_reward = ep_reward
#if test_mode == False:
# multiPG.update_agents()
all_rw.append(ep_reward)
if i_episode % (args.log_interval * 2) == 0 and ifsave_data:
np.save("data/" + model_name + ".npy", all_rw)
if i_episode % args.log_interval == 0:
print(
'Episode {}\tLast reward: {:.2f}\tAverage reward: {:.2f}\ttest_reward: {:.2f}'
.format(i_episode, ep_reward[0], running_reward[0],
test_reward_sum))
logger.scalar_summary("ep_reward", ep_reward, i_episode)
logger.scalar_summary("coin_eaten", test_reward_sum, i_episode)
if i_episode % save_eps == 0 and i_episode > 10 and ifsave_model:
multiPG.save(file_name)