深度学习学习笔记---------RIAL算法/增强智能体间学习

强化学习

DQN算法

马尔可夫决策过程

深度残差网络

多智能体的通信问题

关键公式:

RIAL 算法通信信息也作为一个离散的动作空间来考虑，因此，RIAL需要学习两个Q网络，一个用于输出动作Q值，另一个用于输出通信动作Q值。这里的通信动作是离散的，因此梯度无法从一个智能体传递给另一个智能体。

这段代码用于构建RNN网络，具体为GRU层，输入为opt.model_rnn_size,大小为128，层数为opt.model_rnn_layer=2,若dropout不为0则增加dropout层，之后的连接输出为128的BN层，Relu激活层和一层输入为128输出为opt.game_action_space_toal的全连接层，输出的参数视具体的实验而定。

#RIAL算法/增强智能体间学习

class SwitchCNet(nn,MOdule):
	def __init__(SwitchCNet,self):
		super(SwitchCNet,self).__init__()
		self.opt =opt
		droupt_rate = opt.model_rnn_droupt_rate or 0
		self.rnn = nn.GRU(input_size = opt,model_rnn_size,hidden_size = opt.model_rnn_size,num_layers = opt.model_rnn_layers,droupt = droupt_rate,batch_first = True)

		self.outputs = nn.Sequential()
		if droupt_rate > 0:
			self.outputs.add_moudle('droupt1',nn.Droupt(droupt_rate))
		self.outputs.add_moudle('liner1',nn.liner(opt.model_rnn_size,opt.model_rnn_size))
		if opt.model_bn:
			self.outputs.add_moudle('batchnorm1',nn.batchnorm1d(opt.model_rnn_size))
			self.outputs.add_moudle('relu1',nn.Relu(inplace = True))
			self.outputs.add_moudle('liner2',nn.Linear(opt.model_rnn_size,opt.game_action_space_toal))

o_t为观测值，messages，hidden，prev_action,agent_index为智能体的索引agent_index, o_t , prev_action通过查找表传递，分别为 z_a =self.agent_lookup(agent_index), z_o= self.state_lookup(o_t)和 z_u =self.prev_action_lookup(prev_action).messages通过一个一层MLP，z_m =self.messages_mlp(messages.view(-1,self.comm_size)).输出大小为128，将 z_a, z_o, z_u 和 z_m求和得到z，将Z和内部状态hidden输入RNN网络，输出得到动作outputs和内部状态h_out.

def forward(self,o_t,messages.hidden,prev_action,agent_index):
	opt = self.opt
	o_t = Variable(o_t)
	hidden = Variable(hidden)
	prev_message = None
	if opt.model_dial:
		if opt.model_action_aware:
			prev_action = Variable(prev_action)
		else:
			if opt,model_action_aware:
				prev_action,prev_message = prev_action
				prev_action = Variable(prev_action)
				prev_message = Variable(messages)
		agent_index = Variable(agent_index)
		z_a,z_o,z_u,z_m = [0]*4
		z_a = self.agent_index_lookup(agent_index)
		z_o = self.state_lookup(o_t)
		if opt.model_action_aware:
			z_u = self.prev_action_lookup(prev_action)
			if prev_message is not None:
				z_u += self.prev_message_lookup(prev_message)
		z_m = self.messages_mlp(messages.view(-1,self.comm_size))
		z = z_a+z_o+z_u+z_m
		z = z.unsqueeze(1)
		rnn_out,h_out = self.rnn(z,hidden)
		outputs = self.outputs(rnn_out[:,-1,:].squeeze())
		return h_out,outputs