I'm new in AI, and i want to get in the field, i have spent some time finishing a program to train an agent for a simple customized environment, but when i perform the training in colab for 10000 episodes, it still can not get well performance. I guess whether there is something wrong with the customized env or there is something wrong with the training process.
Env: a helicopter tries to get throw the continous flow of birds (max num: 10), the birds moves from the right to the left, and there is fuel randomly. If the helicopter is still alive, i.e., it has not collided with a bird and still has fuel (initialized by 1000, when it collides with the fuel icon (max num: 2), fuel_left will be reset to 1000), its rewards plus 1.
the environment is shown in the figure:
after 10000 episode in ddpg/dqn, the agent still can not play more than 15 seconds, could you point out where the problem is?
Action space(1 dim): 0, 1, 2, 3, 4 -> helicopter moves up, down, left, right and keep static.
State space(28 dim): (x,y) for 10 birds, 2 fuel, and 1 helicopter. Besides, there is fuel left and rewards obtained.
Rewards: If the helicopter is alive, rewards plus 1.
the env settings code is as follwos (custom.py):
import numpy as np
import cv2
import matplotlib.pyplot as plt
import random
import math
import time
from gym import Env, spaces
import time
font = cv2.FONT_HERSHEY_COMPLEX_SMALL
class ChopperScape(Env):
def __init__(self):
super(ChopperScape,self).__init__()
self.maxbirdnum = 10
self.maxfuelnum = 2
self.observation_shape = (28,)
self.canvas_shape = (600,800,3)
self.action_space = spaces.Discrete(5,)
self.last_action = 0
self.obs = np.zeros(self.observation_shape)
self.canvas = np.ones(self.canvas_shape) * 1
self.elements = []
self.maxfuel = 1000
self.y_min = int (self.canvas_shape[0] * 0.1)
self.x_min = 0
self.y_max = int (self.canvas_shape[0] * 0.9)
self.x_max = self.canvas_shape[1]
def draw_elements_on_canvas(self):
self.canvas = np.ones(self.canvas_shape) * 1
for elem in self.elements:
elem_shape = elem.icon.shape
x,y = elem.x, elem.y
self.canvas[y : y + elem_shape[1], x:x + elem_shape[0]] = elem.icon
text = 'Fuel Left: {} | Rewards: {}'.format(self.fuel_left, self.ep_return)
self.canvas = cv2.putText(self.canvas, text, (10,20), font, 0.8, (0,0,0), 1, cv2.LINE_AA)
def reset(self):
self.fuel_left = self.maxfuel
self.ep_return = 0
self.obs = np.zeros(self.observation_shape)
self.obs[26] = self.maxfuel
self.bird_count = 0
self.fuel_count = 0
x = random.randrange(int(self.canvas_shape[0] * 0.05), int(self.canvas_shape[0] * 0.90))
y = random.randrange(int(self.canvas_shape[1] * 0.05), int(self.canvas_shape[1] * 0.90))
self.chopper = Chopper("chopper", self.x_max, self.x_min, self.y_max, self.y_min)
self.chopper.set_position(x,y)
self.obs[24] = x
self.obs[25] = y
self.elements = [self.chopper]
self.canvas = np.ones(self.canvas_shape) * 1
self.draw_elements_on_canvas()
return self.obs
def get_action_meanings(self):
return {0: "Right", 1: "Left", 2: "Down", 3: "Up", 4: "Do Nothing"}
def has_collided(self, elem1, elem2):
x_col = False
y_col = False
elem1_x, elem1_y = elem1.get_position()
elem2_x, elem2_y = elem2.get_position()
if 2 * abs(elem1_x - elem2_x) <= (elem1.icon_w + elem2.icon_w):
x_col = True
if 2 * abs(elem1_y - elem2_y) <= (elem1.icon_h + elem2.icon_h):
y_col = True
if x_col and y_col:
return True
return False
def step(self, action):
done = False
reward = 1
assert self.action_space.contains(action), "invalid action"
if action == 4:
self.chopper.move(0,5)
elif action == 1:
self.chopper.move(0,-5)
elif action == 2:
self.chopper.move(5,0)
elif action == 0:
self.chopper.move(-5,0)
elif action == 3:
self.chopper.move(0,0)
if random.random() < 0.1 and self.bird_count<self.maxbirdnum:
spawned_bird = Bird("bird_{}".format(self.bird_count), self.x_max, self.x_min, self.y_max, self.y_min)
self.bird_count += 1
bird_y = random.randrange(self.y_min, self.y_max)
spawned_bird.set_position(self.x_max, bird_y)
self.elements.append(spawned_bird)
if random.random() < 0.05 and self.fuel_count<self.maxfuelnum:
spawned_fuel = Fuel("fuel_{}".format(self.bird_count), self.x_max, self.x_min, self.y_max, self.y_min)
self.fuel_count += 1
fuel_x = random.randrange(self.x_min, self.x_max)
fuel_y = self.y_max
spawned_fuel.set_position(fuel_x, fuel_y)
self.elements.append(spawned_fuel)
for elem in self.elements:
if isinstance(elem, Bird):
if elem.get_position()[0] <= self.x_min:
self.elements.remove(elem)
self.bird_count -= 1
else:
elem.move(-5,0)
if self.has_collided(self.chopper, elem):
done = True
reward = -100000.0*(1.0/self.ep_return+1)
if isinstance(elem, Fuel):
flag1 = False
flag2 = False
if self.has_collided(self.chopper, elem):
self.fuel_left = self.maxfuel
flag1 = True
reward += 2
# time.sleep(0.5)
if elem.get_position()[1] <= self.y_min:
flag2 = True
self.fuel_count -= 1
else:
elem.move(0, -5)
if flag1 == True or flag2 == True:
self.elements.remove(elem)
self.fuel_left -= 1
if self.fuel_left == 0:
done = True
self.draw_elements_on_canvas()
self.ep_return += 1
birdnum = 0
fuelnum = 0
x_, y_ = self.chopper.get_position()
dis = 0.0
for elem in self.elements:
x,y = elem.get_position()
if isinstance(elem,Bird):
self.obs[2*birdnum] = x
self.obs[2*birdnum+1] = y
birdnum += 1
dis += math.hypot(x_-x,y_-y)
if isinstance(elem,Fuel):
base = self.maxbirdnum*2
self.obs[base+2*fuelnum] = x
self.obs[base+2*fuelnum+1] = y
fuelnum += 1
self.obs[24] = x_
self.obs[25] = y_
self.obs[26] = self.fuel_left
self.obs[27] = self.ep_return
if x_ == self.x_min or x_ == self.x_max or y_ == self.y_max or y_ == self.y_min:
reward -= random.random()
for i in range(26):
if i%2 == 0:
self.obs[i]/=800.0
else:
self.obs[i]/=600.0
self.obs[26]/=1000.0
self.obs[27]/=100.0
# print('reward:',reward)
# if done == True:
# time.sleep(1)
return self.obs, reward, done, {}
def render(self, mode = "human"):
assert mode in ["human", "rgb_array"], "Invalid mode, must be either \"human\" or \"rgb_array\""
if mode == "human":
cv2.imshow("Game", self.canvas)
cv2.waitKey(10)
elif mode == "rgb_array":
return self.canvas
def close(self):
cv2.destroyAllWindows()
class Point(object):
def __init__(self, name, x_max, x_min, y_max, y_min):
self.x = 0
self.y = 0
self.x_min = x_min
self.x_max = x_max
self.y_min = y_min
self.y_max = y_max
self.name = name
def set_position(self, x, y):
self.x = self.clamp(x, self.x_min, self.x_max - self.icon_w)
self.y = self.clamp(y, self.y_min, self.y_max - self.icon_h)
def get_position(self):
return (self.x, self.y)
def move(self, del_x, del_y):
self.x += del_x
self.y += del_y
self.x = self.clamp(self.x, self.x_min, self.x_max - self.icon_w)
self.y = self.clamp(self.y, self.y_min, self.y_max - self.icon_h)
def clamp(self, n, minn, maxn):
return max(min(maxn, n), minn)
class Chopper(Point):
def __init__(self, name, x_max, x_min, y_max, y_min):
super(Chopper, self).__init__(name, x_max, x_min, y_max, y_min)
self.icon = cv2.imread("chopper1.jpg") / 255.0
self.icon_w = 64
self.icon_h = 64
self.icon = cv2.resize(self.icon, (self.icon_h, self.icon_w))
class Bird(Point):
def __init__(self, name, x_max, x_min, y_max, y_min):
super(Bird, self).__init__(name, x_max, x_min, y_max, y_min)
self.icon = cv2.imread("bird1.jpg") / 255.0
self.icon_w = 32
self.icon_h = 32
self.icon = cv2.resize(self.icon, (self.icon_h, self.icon_w))
class Fuel(Point):
def __init__(self, name, x_max, x_min, y_max, y_min):
super(Fuel, self).__init__(name, x_max, x_min, y_max, y_min)
self.icon = cv2.imread("fuel1.jpg") / 255.0
self.icon_w = 32
self.icon_h = 32
self.icon = cv2.resize(self.icon, (self.icon_h, self.icon_w))
if __name__ == '__main__':
from IPython import display
env = ChopperScape()
obs = env.reset()
while True:
# random agent
action = random.randrange(-1,1)
obs, reward, done, info = env.step(action)
# Render the game
env.render()
if done == True:
break
env.close()
the ddpg algorithm to train the agent is as follows (ddpg.py):
from custom import ChopperScape
import random
import collections
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
#θΆ
εζ°
lr_mu = 0.005
lr_q = 0.01
gamma = 0.99
batch_size = 32
buffer_limit = 50000
tau = 0.005 # for target network soft update
class ReplayBuffer():
def __init__(self):
self.buffer = collections.deque(maxlen=buffer_limit)
def put(self, transition):
self.buffer.append(transition)
def sample(self, n):
mini_batch = random.sample(self.buffer, n)
s_lst, a_lst, r_lst, s_prime_lst, done_mask_lst = [], [], [], [], []
for transition in mini_batch:
s, a, r, s_prime, done = transition
s_lst.append(s)
a_lst.append([a])
r_lst.append(r)
s_prime_lst.append(s_prime)
done_mask = 0.0 if done else 1.0
done_mask_lst.append(done_mask)
return torch.tensor(s_lst, dtype=torch.float), torch.tensor(a_lst, dtype=torch.float), \
torch.tensor(r_lst, dtype=torch.float), torch.tensor(s_prime_lst, dtype=torch.float), \
torch.tensor(done_mask_lst, dtype=torch.float)
def size(self):
return len(self.buffer)
class MuNet(nn.Module):
def __init__(self):
super(MuNet, self).__init__()
self.fc1 = nn.Linear(28, 128)
self.fc2 = nn.Linear(128, 64)
self.fc_mu = nn.Linear(64, 1)
def forward(self, x):
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
mu = torch.tanh(self.fc_mu(x))
return mu
class QNet(nn.Module):
def __init__(self):
super(QNet, self).__init__()
self.fc_s = nn.Linear(28, 64)
self.fc_a = nn.Linear(1,64)
self.fc_q = nn.Linear(128, 32)
self.fc_out = nn.Linear(32,1)
def forward(self, x, a):
h1 = F.relu(self.fc_s(x))
h2 = F.relu(self.fc_a(a))
cat = torch.cat([h1,h2], dim=1)
q = F.relu(self.fc_q(cat))
q = self.fc_out(q)
return q
class OrnsteinUhlenbeckNoise:
def __init__(self, mu):
self.theta, self.dt, self.sigma = 0.1, 0.01, 0.1
self.mu = mu
self.x_prev = np.zeros_like(self.mu)
def __call__(self):
x = self.x_prev + self.theta * (self.mu - self.x_prev) * self.dt + \
self.sigma * np.sqrt(self.dt) * np.random.normal(size=self.mu.shape)
self.x_prev = x
return x
def train(mu, mu_target, q, q_target, memory, q_optimizer, mu_optimizer):
s,a,r,s_prime,done_mask = memory.sample(batch_size)
core = q_target(s_prime, mu_target(s_prime)) * done_mask
target = r + gamma * core
q_loss = F.smooth_l1_loss(q(s,a), target.detach())
q_optimizer.zero_grad()
q_loss.backward()
q_optimizer.step()
mu_loss = -q(s,mu(s)).mean() # That's all for the policy loss.
mu_optimizer.zero_grad()
mu_loss.backward()
mu_optimizer.step()
def soft_update(net, net_target):
for param_target, param in zip(net_target.parameters(), net.parameters()):
param_target.data.copy_(param_target.data * (1.0 - tau) + param.data * tau)
def main():
env = ChopperScape()
memory = ReplayBuffer()
q, q_target = QNet(), QNet()
q_target.load_state_dict(q.state_dict())
mu, mu_target = MuNet(), MuNet()
mu_target.load_state_dict(mu.state_dict())
score = 0.0
print_interval = 20
mu_optimizer = optim.Adam(mu.parameters(), lr=lr_mu)
q_optimizer = optim.Adam(q.parameters(), lr=lr_q)
ou_noise = OrnsteinUhlenbeckNoise(mu=np.zeros(1))
for n_epi in range(10000):
s = env.reset()
done = False
while not done:
a = mu(torch.from_numpy(s).float())
a = a.item() + ou_noise()[0]
print('action:',a)
s_prime, r, done, info = env.step(a)
env.render()
memory.put((s,a,r/100.0,s_prime,done))
score += r
s = s_prime
if memory.size()>20000:
for _ in range(10):
train(mu, mu_target, q, q_target, memory, q_optimizer, mu_optimizer)
soft_update(mu, mu_target)
soft_update(q, q_target)
if n_epi%print_interval==0 and n_epi!=0:
print("# of episode :{}, avg score : {:.1f}".format(n_epi, score/print_interval))
score = 0.0
env.close()
if __name__ == '__main__':
main()
and the dqn algorithm is as follows(dqn.py):
import gym
import collections
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from custom import ChopperScape
#Hyperparameters
learning_rate = 0.0005
gamma = 0.98
buffer_limit = 50000
batch_size = 32
class ReplayBuffer():
def __init__(self):
self.buffer = collections.deque(maxlen=buffer_limit)
def put(self, transition):
self.buffer.append(transition)
def sample(self, n):
mini_batch = random.sample(self.buffer, n)
s_lst, a_lst, r_lst, s_prime_lst, done_mask_lst = [], [], [], [], []
for transition in mini_batch:
s, a, r, s_prime, done_mask = transition
s_lst.append(s)
a_lst.append([a])
r_lst.append([r])
s_prime_lst.append(s_prime)
done_mask_lst.append([done_mask])
return torch.tensor(s_lst, dtype=torch.float), torch.tensor(a_lst), \
torch.tensor(r_lst), torch.tensor(s_prime_lst, dtype=torch.float), \
torch.tensor(done_mask_lst)
def size(self):
return len(self.buffer)
class Qnet(nn.Module):
def __init__(self):
super(Qnet, self).__init__()
self.fc1 = nn.Linear(28, 128)
self.fc2 = nn.Linear(128, 128)
self.fc3 = nn.Linear(128, 5)
def forward(self, x):
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def sample_action(self, obs, epsilon):
out = self.forward(obs)
# coin = random.random()
# if coin < epsilon:
# return random.randint(0,1)
# else :
# return out.argmax().item()
return out.argmax().item()
def train(q, q_target, memory, optimizer):
for _ in range(10):
s,a,r,s_prime,done_mask = memory.sample(batch_size)
q_out = q(s)
q_a = q_out.gather(1,a)
max_q_prime = q_target(s_prime).max(1)[0].unsqueeze(1)
target = r + gamma * max_q_prime * done_mask
loss = F.smooth_l1_loss(q_a, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
def main():
env = ChopperScape()
q = torch.load('10000_dqn_3.pt')
q_target = torch.load('10000_dqn_3_qtarget.pt')
# q_target.load_state_dict(q.state_dict())
memory = ReplayBuffer()
print_interval = 20
score = 0.0
optimizer = optim.Adam(q.parameters(), lr=learning_rate)
for n_epi in range(10000):
epsilon = max(0.01, 0.08 - 0.01*(n_epi/200)) #Linear annealing from 8% to 1%
s = env.reset()
done = False
while not done:
a = q.sample_action(torch.from_numpy(s).float(), epsilon)
s_prime, r, done, info = env.step(a)
env.render()
done_mask = 0.0 if done else 1.0
memory.put((s,a,r,s_prime, done_mask))
s = s_prime
if done:
break
score += r
if memory.size()>20000:
train(q, q_target, memory, optimizer)
if n_epi%print_interval==0 and n_epi!=0:
q_target.load_state_dict(q.state_dict())
print("n_episode :{}, score : {:.1f}, n_buffer : {}, eps : {:.1f}%".format(n_epi, score/print_interval, memory.size(), epsilon*100))
score = 0.0
env.close()
def test():
env = ChopperScape()
q = torch.load('10000_dqn_q.pt')
done = False
s = env.reset()
while not done:
a = q.sample_action(torch.from_numpy(s).float(), 1)
s_prime, r, done, info = env.step(a)
env.render()
s = s_prime
if done:
break
if __name__ == '__main__':
main()
when perform dqn, please annotate the action convert part in custom.py/class ChoperScape/step
after 10000 episode in ddpg/dqn, the agent still can not play more than 15 seconds, could you point out where the problem is?
