TaskAllocation/manager.py at main · yk7333/TaskAllocation · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
import numpy as np
import torch as th
import torch.nn as nn
import itertools
import collections
import random

class ManagerReplayBuffer:
    def __init__(self, capacity):
        self.buffer = collections.deque(maxlen=capacity)

    def add(self, state, agent_type, action, reward, next_state, done,avail_agents):
        self.buffer.append((state, agent_type, action, reward, next_state, done, avail_agents))

    def sample(self, batch_size):
        transitions = random.sample(self.buffer, batch_size)
        state, agent_type, action, reward, next_state, done, avail_agents = zip(*transitions)
        return np.array(state), np.array(agent_type), action, reward, np.array(next_state), done, avail_agents

    def size(self):
        return len(self.buffer)

class Embedding(nn.Module):
    def __init__(self,input_dim,hidden_dim):
        super(Embedding,self).__init__()
        self.fc1 = nn.Linear(input_dim,hidden_dim)
        self.fc2 = nn.Linear(hidden_dim,hidden_dim)
        self.fc4 = nn.Linear(hidden_dim,hidden_dim)
        self.relu = th.relu
        self.hidden_dim = hidden_dim

    def forward(self,x):
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        return x.reshape(-1,self.hidden_dim)

class ManagerNet(nn.Module):
    def __init__(self,agent_dim,task_dim,hidden_dim,agent_type,device):
        super(ManagerNet,self).__init__()
        self.agent_type = th.tensor(agent_type,dtype=th.float,device=device).squeeze(0)
        self.agent_embed = Embedding(agent_dim,hidden_dim)
        self.task_embed = Embedding(task_dim,hidden_dim)
        self.fc_task = nn.Sequential(nn.Linear(task_dim,hidden_dim),nn.ReLU(),nn.Linear(hidden_dim,hidden_dim))
        self.fc_agent = nn.Sequential(nn.Linear(agent_dim,hidden_dim),nn.ReLU(),nn.Linear(hidden_dim,hidden_dim))
        self.hidden_dim = hidden_dim
        self.device = device

    def reset_agent(self,agent_type):
        self.agent_type = th.tensor(agent_type,dtype=th.float,device=self.device).squeeze(0)

    def actor(self,task,avail_action):
        avail_action = th.tensor(avail_action,dtype=th.float,device=self.device) #batch_size, num_agents
        if len(avail_action.shape)==1:
            avail_action = avail_action.unsqueeze(0)
        task = th.tensor(task,dtype=th.float,device=self.device)    #batch_size, task_dim
        agent_embedding = self.agent_embed(self.agent_type) #num_agents,hidden_dim
        task_embedding = self.task_embed(task)  #batch_size, hidden_dim
        score = agent_embedding @ task_embedding.T /th.tensor(self.hidden_dim,dtype=th.long,device=self.device)    #num_agents,batch_size
        probs = th.softmax(th.where(avail_action.T==1,score,-999999.),dim=0)    #num_agents,batch_size
        return probs.T #batch_size,num_agents

    def critic(self,task):
        Q_value = self.fc_agent(self.agent_type) @ self.fc_task(task).reshape(-1,1)  #agent_num,hidden_dim @ hidden_dim,1 -> agent_num,1
        return Q_value.T


class Manager:
    def __init__(self,state_dim,agent_dim,hidden_dim,actor_lr,critic_lr,alpha,tau,gamma,device,agent_type,model_dir):
        self.net = ManagerNet(agent_dim,state_dim,hidden_dim,agent_type,device).to(device)
        self.target_net = ManagerNet(agent_dim,state_dim,hidden_dim,agent_type,device).to(device)
        self.target_net.load_state_dict(self.net.state_dict())
        self.hidden_dim = hidden_dim
        self.state_dim = state_dim
        self.loss = nn.MSELoss()
        actor_params = itertools.chain(self.net.agent_embed.parameters(), self.net.task_embed.parameters())
        critic_params = itertools.chain(self.net.fc_task.parameters(), self.net.fc_agent.parameters())
        self.actor_optimizer = th.optim.Adam(actor_params,lr=actor_lr)
        self.critic_optimizer = th.optim.Adam(critic_params,lr=critic_lr)
        self.alpha = alpha
        self.tau = tau
        self.gamma = gamma
        self.device=device
        self.critic_loss_list = []
        self.actor_loss_list = []
        self.model_dir = model_dir

    def reset_agent(self,agent_type):
        self.net.reset_agent(agent_type)
        self.target_net.reset_agent(agent_type)

    def select(self,state, avail_action,evaluate=False):
        probs = self.net.actor(state,avail_action)
        if evaluate:
            action = probs.argmax()
        else:
            action_dist = th.distributions.Categorical(probs)
            action = action_dist.sample()
        return action.cpu().detach().item()

    def update(self,transition_dict):
        states = th.tensor(transition_dict['states'],dtype=th.float).flatten().to(self.device)
        agent_type = th.tensor(transition_dict['agent_type'],dtype=th.float).to(self.device)
        actions = th.tensor(transition_dict['actions'],dtype=th.long).to(self.device)
        rewards = th.tensor(transition_dict['rewards'],dtype=th.float).to(self.device)
        next_states = th.tensor(transition_dict['next_states'],dtype=th.float).flatten().to(self.device)
        dones = th.tensor(transition_dict['dones'],dtype=th.float).to(self.device)
        avail_agents = th.tensor(transition_dict['avail_agents'],dtype=th.float).flatten().to(self.device)

        self.net.reset_agent(agent_type)
        self.target_net.reset_agent(agent_type)
        next_probs = self.net.actor(next_states, avail_agents)
        next_log_probs = th.log(next_probs + 1e-8)
        entropy = -th.sum(next_probs * next_log_probs, dim=1, keepdim=True)
        q_value = self.net.critic(next_states)
        v_value = th.sum(next_probs * q_value,dim=1,keepdim=True).flatten()
        next_value = v_value + 0.01 * entropy
        td_target = rewards + self.gamma * next_value * (1 - dones)
        critic_q_values = self.net.critic(states).reshape(1,-1).gather(1, actions.reshape(1,1))
        critic_loss = self.loss(td_target,critic_q_values)
        self.critic_loss_list.append(critic_loss.cpu().detach().item())
        self.critic_optimizer.zero_grad()
        critic_loss.backward()
        self.critic_optimizer.step()

        avail_agents.scatter_(0,actions,1)
        probs = self.net.actor(states,avail_agents)
        log_probs = th.log(probs + 1e-8)
        entropy = -th.sum(probs * log_probs, dim=1, keepdim=True)
        q_value = self.target_net.critic(states)
        v_value = th.sum(probs * q_value,dim=1,keepdim=True)
        actor_loss = th.mean(-0.01 * entropy - v_value)
        self.actor_loss_list.append(actor_loss.cpu().detach().item())
        self.actor_optimizer.zero_grad()
        actor_loss.backward()
        self.actor_optimizer.step()
        self.soft_update(self.net, self.target_net)

    def soft_update(self, net, target_net):
        for param_target, param in zip(target_net.parameters(), net.parameters()):
            param_target.data.copy_(param_target.data * (1.0 - self.tau) + param.data * self.tau)


    def save(self):
        model_path = self.model_dir + "/model.pt"
        th.save(self.net.state_dict(),model_path)

    def load(self):
        model_path = self.model_dir + "/model.pt"
        self.net.load_state_dict(th.load(model_path))
        self.target_net.load_state_dict(th.load(model_path))