update

codes/HierarchicalDQN/README.md

@@ -0,0 +1,13 @@
+# Hierarchical DQN
+
+## 原理简介
+
+Hierarchical DQN是一种分层强化学习方法，与DQN相比增加了一个meta controller，
+
+![image-20210331153115575](assets/image-20210331153115575.png)
+
+即学习时，meta controller每次会生成一个goal，然后controller或者说下面的actor就会达到这个goal，直到done为止。这就相当于给agent增加了一个队长，队长擅长制定局部目标，指导agent前行，这样应对一些每回合步数较长或者稀疏奖励的问题会有所帮助。
+
+## 伪代码
+
+![image-20210331153542314](assets/image-20210331153542314.png)

codes/HierarchicalDQN/agent.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-24 22:18:18
 LastEditor: John
-LastEditTime: 2021-03-27 04:24:30
+LastEditTime: 2021-03-31 14:51:09
 Discription: 
 Environment: 
 '''
@@ -13,90 +13,103 @@ import torch
 import torch.nn as nn
 import numpy as np
 import random,math
-from HierarchicalDQN.model import MLP
-from common.memory import ReplayBuffer
 import torch.optim as optim
+from common.model import MLP
+from common.memory import ReplayBuffer
+
 class HierarchicalDQN:
     def __init__(self,state_dim,action_dim,cfg):
+        self.state_dim = state_dim
         self.action_dim = action_dim
+        self.gamma = cfg.gamma
         self.device = cfg.device
         self.batch_size = cfg.batch_size
-        self.sample_count = 0 
-        self.epsilon = 0
-        self.epsilon_start = cfg.epsilon_start
-        self.epsilon_end = cfg.epsilon_end
-        self.epsilon_decay = cfg.epsilon_decay
-        self.batch_size = cfg.batch_size
+        self.frame_idx = 0 
+        self.epsilon = lambda frame_idx: cfg.epsilon_end + (cfg.epsilon_start - cfg.epsilon_end ) * math.exp(-1. * frame_idx / cfg.epsilon_decay)
         self.policy_net = MLP(2*state_dim, action_dim,cfg.hidden_dim).to(self.device)
-        self.target_net = MLP(2*state_dim, action_dim,cfg.hidden_dim).to(self.device)
-        self.meta_policy_net  = MLP(state_dim, state_dim,cfg.hidden_dim).to(self.device)
-        self.meta_target_net = MLP(state_dim, state_dim,cfg.hidden_dim).to(self.device)
+        self.meta_policy_net = MLP(state_dim, state_dim,cfg.hidden_dim).to(self.device)
         self.optimizer = optim.Adam(self.policy_net.parameters(),lr=cfg.lr)
         self.meta_optimizer = optim.Adam(self.meta_policy_net.parameters(),lr=cfg.lr)
         self.memory = ReplayBuffer(cfg.memory_capacity)
         self.meta_memory = ReplayBuffer(cfg.memory_capacity)
-    def to_onehot(x):
-        oh = np.zeros(6)
+        self.loss_numpy  = 0
+        self.meta_loss_numpy  = 0
+        self.losses = []
+        self.meta_losses = []
+    def to_onehot(self,x):
+        oh = np.zeros(self.state_dim)
         oh[x - 1] = 1.
         return oh
-    def set_goal(self,meta_state):
-        self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * math.exp(-1. * self.sample_count / self.epsilon_decay)
-        self.sample_count += 1
-        if random.random() > self.epsilon:
+    def set_goal(self,state):
+        if random.random() > self.epsilon(self.frame_idx):
             with torch.no_grad():
-                meta_state = torch.tensor([meta_state], device=self.device, dtype=torch.float32)
-                q_value = self.policy_net(meta_state)
-                goal = q_value.max(1)[1].item() 
+                state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(0)
+                goal = self.meta_policy_net(state).max(1)[1].item() 
         else:
-            goal = random.randrange(self.action_dim)
-        goal = self.meta_policy_net(meta_state)
-        onehot_goal = self.to_onehot(goal)
-        return onehot_goal
+            goal = random.randrange(self.state_dim)
+        return goal
     def choose_action(self,state):
-        self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * math.exp(-1. * self.sample_count / self.epsilon_decay)
-        self.sample_count += 1
-        if random.random() > self.epsilon:
+        self.frame_idx += 1
+        if random.random() > self.epsilon(self.frame_idx):
             with torch.no_grad():
-                state = torch.tensor([state], device=self.device, dtype=torch.float32)
+                state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(0)
                 q_value = self.policy_net(state)
                 action = q_value.max(1)[1].item()  
         else:
             action = random.randrange(self.action_dim)
         return action
     def update(self):
+        self.update_policy()
+        self.update_meta()
+    def update_policy(self): 
         if self.batch_size > len(self.memory):
-            state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(self.batch_size)
-        state_batch = torch.tensor(
-            state_batch, device=self.device, dtype=torch.float)
-        action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(1)  
-        reward_batch = torch.tensor(reward_batch, device=self.device, dtype=torch.float)  
-        next_state_batch = torch.tensor(next_state_batch, device=self.device, dtype=torch.float)
-        done_batch = torch.tensor(np.float32(done_batch), device=self.device).unsqueeze(1)  
-        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch)
-        next_state_values = self.target_net(next_state_batch).max(1)[0].detach()  
-        expected_q_values = reward_batch + self.gamma * next_state_values * (1-done_batch[0])
-        loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1)) 
+            return
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(self.batch_size)
+        state_batch = torch.tensor(state_batch,dtype=torch.float)
+        action_batch = torch.tensor(action_batch,dtype=torch.int64).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch,dtype=torch.float)  
+        next_state_batch = torch.tensor(next_state_batch, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch))
+        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch).squeeze(1)
+        next_state_values = self.policy_net(next_state_batch).max(1)[0].detach()
+        expected_q_values = reward_batch + 0.99 * next_state_values * (1-done_batch)
+        loss = nn.MSELoss()(q_values, expected_q_values) 
         self.optimizer.zero_grad() 
         loss.backward()
-        for param in self.policy_net.parameters(): 
+        for param in self.policy_net.parameters():  # clip防止梯度爆炸
             param.grad.data.clamp_(-1, 1)
-        self.optimizer.step() 
-
+        self.optimizer.step()  
+        self.loss_numpy = loss.detach().numpy()
+        self.losses.append(self.loss_numpy)  
+    def update_meta(self):
         if self.batch_size > len(self.meta_memory):
-            meta_state_batch, meta_action_batch, meta_reward_batch, next_meta_state_batch, meta_done_batch = self.memory.sample(self.batch_size)
-        meta_state_batch = torch.tensor(meta_state_batch, device=self.device, dtype=torch.float)
-        meta_action_batch = torch.tensor(meta_action_batch, device=self.device).unsqueeze(1)  
-        meta_reward_batch = torch.tensor(meta_reward_batch, device=self.device, dtype=torch.float)  
-        next_meta_state_batch = torch.tensor(next_meta_state_batch, device=self.device, dtype=torch.float)
-        meta_done_batch = torch.tensor(np.float32(meta_done_batch), device=self.device).unsqueeze(1)  
-        meta_q_values = self.meta_policy_net(meta_state_batch).gather(dim=1, index=meta_action_batch)
-        next_state_values = self.target_net(next_meta_state_batch).max(1)[0].detach()  
-        expected_meta_q_values = meta_reward_batch + self.gamma * next_state_values * (1-meta_done_batch[0])
-        meta_loss = nn.MSEmeta_loss()(meta_q_values, expected_meta_q_values.unsqueeze(1)) 
+            return
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.meta_memory.sample(self.batch_size)
+        state_batch = torch.tensor(state_batch,dtype=torch.float)
+        action_batch = torch.tensor(action_batch,dtype=torch.int64).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch,dtype=torch.float)  
+        next_state_batch = torch.tensor(next_state_batch, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch))
+        q_values = self.meta_policy_net(state_batch).gather(dim=1, index=action_batch).squeeze(1)
+        next_state_values = self.meta_policy_net(next_state_batch).max(1)[0].detach()
+        expected_q_values = reward_batch + 0.99 * next_state_values * (1-done_batch)
+        meta_loss = nn.MSELoss()(q_values, expected_q_values) 
         self.meta_optimizer.zero_grad() 
         meta_loss.backward()
-        for param in self.meta_policy_net.parameters(): 
+        for param in self.meta_policy_net.parameters():  # clip防止梯度爆炸
             param.grad.data.clamp_(-1, 1)
         self.meta_optimizer.step() 
+        self.meta_loss_numpy = meta_loss.detach().numpy()
+        self.meta_losses.append(self.meta_loss_numpy)
+
+    def save(self, path):
+        torch.save(self.policy_net.state_dict(), path+'policy_checkpoint.pth')
+        torch.save(self.meta_policy_net.state_dict(), path+'meta_checkpoint.pth')
+
+    def load(self, path):
+        self.policy_net.load_state_dict(torch.load(path+'policy_checkpoint.pth'))
+        self.meta_policy_net.load_state_dict(torch.load(path+'meta_checkpoint.pth'))
+        
+
         
         

codes/HierarchicalDQN/main.py

@@ -3,95 +3,108 @@
 '''
 Author: John
 Email: johnjim0816@gmail.com
-Date: 2021-03-24 22:14:04
+Date: 2021-03-29 10:37:32
 LastEditor: John
-LastEditTime: 2021-03-27 04:23:43
+LastEditTime: 2021-03-31 14:58:49
 Discription: 
 Environment: 
 '''
+
+
 import sys,os
-sys.path.append(os.getcwd()) # add current terminal path to sys.path
-import gym
+curr_path = os.path.dirname(__file__)
+parent_path = os.path.dirname(curr_path)
+sys.path.append(parent_path)  # add current terminal path to sys.path
+
+import datetime
 import numpy as np
 import torch
-import datetime
-from HierarchicalDQN.agent import HierarchicalDQN
-from common.plot import plot_rewards
-from common.utils import save_results
+import gym
 
-SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
-SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/'  # path to save model
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"): 
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
+from common.utils import save_results
+from common.plot import plot_rewards,plot_losses
+from HierarchicalDQN.agent import HierarchicalDQN
+
+SEQUENCE = datetime.datetime.now().strftime(
+    "%Y%m%d-%H%M%S")  # obtain current time
+SAVED_MODEL_PATH = curr_path+"/saved_model/"+SEQUENCE+'/'  # path to save model
+if not os.path.exists(curr_path+"/saved_model/"):
+    os.mkdir(curr_path+"/saved_model/")
 if not os.path.exists(SAVED_MODEL_PATH):
     os.mkdir(SAVED_MODEL_PATH)
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/' # path to save rewards
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/results/"): 
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
-if not os.path.exists(RESULT_PATH): 
+RESULT_PATH = curr_path+"/results/"+SEQUENCE+'/'  # path to save rewards
+if not os.path.exists(curr_path+"/results/"):
+    os.mkdir(curr_path+"/results/")
+if not os.path.exists(RESULT_PATH):
     os.mkdir(RESULT_PATH)
 
+
 class HierarchicalDQNConfig:
     def __init__(self):
-        self.algo = "DQN" # name of algo
+        self.algo = "H-DQN"  # name of algo
         self.gamma = 0.99
-        self.epsilon_start = 0.95 # start epsilon of e-greedy policy
+        self.epsilon_start = 1  # start epsilon of e-greedy policy
         self.epsilon_end = 0.01
         self.epsilon_decay = 200
-        self.lr = 0.01 # learning rate
-        self.memory_capacity = 800 # Replay Memory capacity
-        self.batch_size = 64
-        self.train_eps = 250 # 训练的episode数目
-        self.train_steps = 200 # 训练每个episode的最大长度
-        self.target_update = 2 # target net的更新频率
-        self.eval_eps = 20 # 测试的episode数目
-        self.eval_steps = 200 # 测试每个episode的最大长度
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
-        self.hidden_dim = 256 # dimension of hidden layer
+        self.lr = 0.0001  # learning rate
+        self.memory_capacity = 10000  # Replay Memory capacity
+        self.batch_size = 32
+        self.train_eps = 300  # 训练的episode数目
+        self.target_update = 2  # target net的更新频率
+        self.eval_eps = 20  # 测试的episode数目
+        self.device = torch.device(
+            "cuda" if torch.cuda.is_available() else "cpu")  # 检测gpu
+        self.hidden_dim = 256  # dimension of hidden layer
 
-def train(cfg,env,agent):
+
+def train(cfg, env, agent):
     print('Start to train !')
     rewards = []
-    ma_rewards = [] # moving average reward
-    ep_steps = []
+    ma_rewards = []  # moveing average reward
     for i_episode in range(cfg.train_eps):
-        state = env.reset() 
-        extrinsic_reward = 0
-        for i_step in range(cfg.train_steps):
-            goal= agent.set_goal(state)
+        state = env.reset()
+        done = False
+        ep_reward = 0
+        while not done:
+            goal = agent.set_goal(state)
+            onehot_goal = agent.to_onehot(goal)
             meta_state = state
-            goal_state  = np.concatenate([state, goal])
-            action = agent.choose_action(state) 
-            next_state, reward, done, _ = env.step(action)
-            extrinsic_reward += reward
-            intrinsic_reward = 1.0 if goal == np.argmax(next_state) else 0.0
-            agent.memory.push(goal_state, action, intrinsic_reward, np.concatenate([next_state, goal]), done)
-            state = next_state 
-            agent.update()
-            if done:
-                break
-        if i_episode % cfg.target_update == 0:
-            agent.target_net.load_state_dict(agent.policy_net.state_dict())
-        print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,extrinsic_reward,i_step+1,done))
-        ep_steps.append(i_step)
-        rewards.append(extrinsic_reward)
+            extrinsic_reward = 0
+            while not done and goal != np.argmax(state):
+                goal_state = np.concatenate([state, onehot_goal])
+                action = agent.choose_action(goal_state)
+                next_state, reward, done, _ = env.step(action)
+                ep_reward += reward
+                extrinsic_reward += reward
+                intrinsic_reward = 1.0 if goal == np.argmax(
+                    next_state) else 0.0
+                agent.memory.push(goal_state, action, intrinsic_reward, np.concatenate(
+                    [next_state, onehot_goal]), done)
+                state = next_state
+                agent.update()
+        agent.meta_memory.push(meta_state, goal, extrinsic_reward, state, done)
+        print('Episode:{}/{}, Reward:{}, Loss:{:.2f}, Meta_Loss:{:.2f}'.format(i_episode+1, cfg.train_eps, ep_reward,agent.loss_numpy ,agent.meta_loss_numpy ))
+        rewards.append(ep_reward)
         if ma_rewards:
             ma_rewards.append(
-                0.9*ma_rewards[-1]+0.1*extrinsic_reward)
+                0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
-            ma_rewards.append(extrinsic_reward)   
-    agent.meta_memory.push(meta_state, goal, extrinsic_reward, state, done)
+            ma_rewards.append(ep_reward)
     print('Complete training！')
-    return rewards,ma_rewards
+    return rewards, ma_rewards
+
 
 if __name__ == "__main__":
-    cfg = HierarchicalDQNConfig()
     env = gym.make('CartPole-v0')
-    env.seed(1) 
+    env.seed(1)
+    cfg = HierarchicalDQNConfig()
     state_dim = env.observation_space.shape[0]
     action_dim = env.action_space.n
-    agent = HierarchicalDQN(state_dim,action_dim,cfg)
-    rewards,ma_rewards = train(cfg,env,agent)
+    agent = HierarchicalDQN(state_dim, action_dim, cfg)
+    rewards, ma_rewards = train(cfg, env, agent)
     agent.save(path=SAVED_MODEL_PATH)
-    save_results(rewards,ma_rewards,tag='train',path=RESULT_PATH)
-    plot_rewards(rewards,ma_rewards,tag="train",algo = cfg.algo,path=RESULT_PATH)
+    save_results(rewards, ma_rewards, tag='train', path=RESULT_PATH)
+    plot_rewards(rewards, ma_rewards, tag="train",
+                 algo=cfg.algo, path=RESULT_PATH)
+    plot_losses(agent.losses,algo=cfg.algo, path=RESULT_PATH)
+

codes/HierarchicalDQN/model.py

@@ -1,24 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2021-03-24 22:14:12
-LastEditor: John
-LastEditTime: 2021-03-24 22:17:09
-Discription: 
-Environment: 
-'''
-import torch.nn as nn
-import torch.nn.functional as F
-class MLP(nn.Module):
-    def __init__(self, state_dim,action_dim,hidden_dim=128):
-        super(MLP, self).__init__()
-        self.fc1 = nn.Linear(state_dim, hidden_dim) 
-        self.fc2 = nn.Linear(hidden_dim,hidden_dim)
-        self.fc3 = nn.Linear(hidden_dim, action_dim) 
-        
-    def forward(self, x):
-        x = F.relu(self.fc1(x)) 
-        x = F.relu(self.fc2(x))
-        return self.fc3(x)