def __init__(self, args):
self.n_actions = args.n_actions
self.n_agents = args.n_agents
self.state_shape = args.state_shape
self.obs_shape = args.obs_shape
alg = args.alg
if alg.find('reinforce') > -1:
self.policy = Reinforce(args)
elif alg.find('coma') > -1:
self.policy = COMA(args)
elif alg.find('central_v') > -1:
self.policy = CentralV(args)
else:
raise Exception("No such algorithm")
self.args = args
print('Init CommAgents')
def __init__(self, args):
self.n_actions = args.n_actions
self.n_agents = args.n_agents
self.state_shape = args.state_shape
self.obs_shape = args.obs_shape
if args.alg == 'vdn':
from policy.vdn import VDN
self.policy = VDN(args)
elif args.alg == 'iql':
from policy.iql import IQL
self.policy = IQL(args)
elif args.alg == 'qmix':
from policy.qmix import QMIX
self.policy = QMIX(args)
elif args.alg == 'coma':
from policy.coma import COMA
self.policy = COMA(args)
elif args.alg == 'qtran_alt':
from policy.qtran_alt import QtranAlt
self.policy = QtranAlt(args)
elif args.alg == 'qtran_base':
from policy.qtran_base import QtranBase
self.policy = QtranBase(args)
elif args.alg == 'maven':
from policy.maven import MAVEN
self.policy = MAVEN(args)
elif args.alg == 'central_v':
from policy.central_v import CentralV
self.policy = CentralV(args)
elif args.alg == 'reinforce':
from policy.reinforce import Reinforce
self.policy = Reinforce(args)
else:
raise Exception("No such algorithm")
self.args = args
def __init__(self, args):
self.n_actions = args.n_actions
self.n_agents = args.n_agents
self.state_shape = args.state_shape
self.obs_shape = args.obs_shape
if args.alg == 'vdn':
self.policy = VDN(args)
elif args.alg == 'qmix':
self.policy = QMIX(args)
elif args.alg == 'coma':
self.policy = COMA(args)
elif args.alg == 'qtran_alt':
self.policy = QtranAlt(args)
elif args.alg == 'qtran_base':
self.policy = QtranBase(args)
elif args.alg == 'maven':
self.policy = MAVEN(args)
elif args.alg == 'central_v':
self.policy = CentralV(args)
elif args.alg == 'reinforce':
self.policy = Reinforce(args)
else:
raise Exception("No such algorithm")
self.args = args
print('Init Agents')
def __init__(self, args):
self.n_actions = args.n_actions
self.n_agents = args.n_agents * 2
self.state_shape = args.state_shape
self.obs_shape = args.obs_shape
self.idact_shape = args.id_dim + args.n_actions
self.search_actions = np.eye(args.n_actions)
self.search_ids = np.zeros(self.n_agents)
if args.alg == 'vdn':
self.policy = VDN(args)
elif args.alg == 'qmix':
self.policy = QMIX(args)
elif args.alg == 'ours':
self.policy = OURS(args)
elif args.alg == 'coma':
self.policy = COMA(args)
elif args.alg == 'qtran_alt':
self.policy = QtranAlt(args)
elif args.alg == 'qtran_base':
self.policy = QtranBase(args)
elif args.alg == 'maven':
self.policy = MAVEN(args)
elif args.alg == 'central_v':
self.policy = CentralV(args)
elif args.alg == 'reinforce':
self.policy = Reinforce(args)
else:
raise Exception("No such algorithm")
if args.use_fixed_model:
args_goal_a = get_common_args()
args_goal_a.load_model = True
args_goal_a = get_mixer_args(args_goal_a)
args_goal_a.learn = False
args_goal_a.epsilon = 0 # 1
args_goal_a.min_epsilon = 0
args_goal_a.map = 'battle'
args_goal_a.n_actions = args.n_actions
args_goal_a.episode_limit = args.episode_limit
args_goal_a.n_agents = args.n_agents
args_goal_a.state_shape = args.state_shape
args_goal_a.feature_shape = args.feature_shape
args_goal_a.view_shape = args.view_shape
args_goal_a.obs_shape = args.obs_shape
args_goal_a.real_view_shape = args.real_view_shape
args_goal_a.load_num = args.load_num
args_goal_a.use_ja = False
args_goal_a.mlp_hidden_dim = [512, 512]
self.fixed_policy = VDN_F(args_goal_a)
self.args = args
print('Init Agents')
class CommAgents:
def __init__(self, args):
self.n_actions = args.n_actions
self.n_agents = args.n_agents
self.state_shape = args.state_shape
self.obs_shape = args.obs_shape
alg = args.alg
if alg.find('reinforce') > -1:
self.policy = Reinforce(args)
elif alg.find('coma') > -1:
self.policy = COMA(args)
elif alg.find('central_v') > -1:
self.policy = CentralV(args)
else:
raise Exception("No such algorithm")
self.args = args
print('Init CommAgents')
# 根据weights得到概率，然后再根据epsilon选动作
def choose_action(self, weights, avail_actions, epsilon, evaluate=False):
weights = weights.unsqueeze(0)
avail_actions = torch.tensor(avail_actions,
dtype=torch.float32).unsqueeze(0)
action_num = avail_actions.sum(dim=1, keepdim=True).float().repeat(
1, avail_actions.shape[-1]) # 可以选择的动作的个数
# 先将Actor网络的输出通过softmax转换成概率分布
prob = torch.nn.functional.softmax(weights, dim=-1)
# 在训练的时候给概率分布添加噪音
prob = ((1 - epsilon) * prob +
torch.ones_like(prob) * epsilon / action_num)
prob[avail_actions == 0] = 0.0 # 不能执行的动作概率为0
"""
不能执行的动作概率为0之后，prob中的概率和不为1，这里不需要进行正则化，因为torch.distributions.Categorical
会将其进行正则化。要注意在训练的过程中没有用到Categorical，所以训练时取执行的动作对应的概率需要再正则化。
"""
if epsilon == 0 and evaluate:
# 测试时直接选最大的
action = torch.argmax(prob)
else:
action = Categorical(prob).sample().long()
return action
def get_action_weights(self, obs, last_action):
obs = torch.tensor(obs, dtype=torch.float32)
last_action = torch.tensor(last_action, dtype=torch.float32)
inputs = list()
inputs.append(obs)
# 给obs添加上一个动作、agent编号
if self.args.last_action:
inputs.append(last_action)
if self.args.reuse_network:
inputs.append(torch.eye(self.args.n_agents))
inputs = torch.cat([x for x in inputs], dim=1)
if self.args.cuda:
inputs = inputs.cuda()
self.policy.eval_hidden = self.policy.eval_hidden.cuda()
weights, self.policy.eval_hidden = self.policy.eval_rnn(
inputs, self.policy.eval_hidden)
weights = weights.reshape(self.args.n_agents, self.args.n_actions)
return weights.cpu()
def _get_max_episode_len(self, batch):
terminated = batch['terminated']
episode_num = terminated.shape[0]
max_episode_len = 0
for episode_idx in range(episode_num):
for transition_idx in range(self.args.episode_limit):
if terminated[episode_idx, transition_idx, 0] == 1:
if transition_idx + 1 >= max_episode_len:
max_episode_len = transition_idx + 1
break
return max_episode_len
def train(self,
batch,
train_step,
epsilon=None): # coma在训练时也需要epsilon计算动作的执行概率
# 每次学习时，各个episode的长度不一样，因此取其中最长的episode作为所有episode的长度
max_episode_len = self._get_max_episode_len(batch)
for key in batch.keys():
batch[key] = batch[key][:, :max_episode_len]
self.policy.learn(batch, max_episode_len, train_step, epsilon)
if train_step > 0 and train_step % self.args.save_cycle == 0:
self.policy.save_model(train_step)