def main():
# Init env
env = gym.make(env_name)
# l_obs = env.observation_space.shape[0]
l_obs = 1
n_action = env.action_space.n
epsilon = 0.5
epsilon_end = 0.01
epsilon_div = 0.025
epsilon_step = ((epsilon - epsilon_end) / epsilon_div)
# Init net model
net = ConvNet(l1, l2, l_obs, n_action)
net.apply(model.weight_init)
optimizer = torch.optim.Adam(net.parameters(), lr=lr)
# Iteration start
for iteration in range(num_iter):
buff = []
print("iterations: ", iteration + 1, "/ ", num_iter)
# Epoch start
for epoch in range(num_epoch):
obs = env.reset()
next_obs = None
reward = 0
total_reward = 0
done = False
traj = Trajectory()
while not done:
if next_obs is None:
obs = process_frame(obs)
obs = torch.Tensor(obs).unsqueeze(0)
action = action_decide(net, obs, n_action, epsilon)
next_obs, reward, done, info = env.step(action)
next_obs = process_frame(next_obs)
next_obs = torch.Tensor(next_obs).unsqueeze(0)
traj.add(obs, action, reward, next_obs)
total_reward += reward
obs = next_obs
if iteration % 10 == 0 and epoch == 1:
env.render()
if done:
buff.append(traj)
reward_list.append(total_reward)
print('reward: ', total_reward)
A2C(net, optimizer, buff)
if epsilon > epsilon_end:
epsilon -= epsilon_step
else:
epsilon = epsilon_end
env.close()
return reward_list
def main():
# Init env
env = gym.make(env_name)
l_obs = 4
n_action = env.action_space.n
# Init net model
actor = Net(num_fc_a, num_fc_a, l_obs, n_action)
critic = Net(num_fc_c, num_fc_c, l_obs, 1)
actor.apply(model.weight_init)
critic.apply(model.weight_init)
optimizer_a = torch.optim.Adam(actor.parameters(), lr=0.01)
optimizer_c = torch.optim.Adam(critic.parameters(), lr=0.01)
# Iteration start
for iteration in range(num_iter):
buff = []
# Epoch start
for epoch in range(num_epoch):
obs = env.reset()
next_obs = None
reward = 0
total_reward = 0
done = False
traj = Trajectory()
while not done:
if next_obs is not None:
obs = next_obs
obs = torch.Tensor(obs).float()
action = action_decide(actor, obs)
next_obs, reward, done, info = env.step(action)
next_obs = torch.tensor(next_obs).float()
traj.add(obs, action, reward, next_obs)
total_reward += reward
# if epoch == 1:
# env.render()
if done:
buff.append(traj)
reward_list.append(total_reward)
A2C(actor, critic, optimizer_a, optimizer_c, buff)
env.close()
return reward_list
def main():
# Init env
env = gym.make(env_name)
l_obs = env.observation_space.shape[0]
n_action = env.action_space.n
# Init net model
net = Net_LSTM(num_fc_a, num_fc_a, l_obs, n_action)
net.apply(model.weight_init)
# optimizer = torch.optim.Adam(net.parameters(), lr=0.01)
# Iteration start
for iteration in range(num_iter):
buff = []
# Epoch start
for epoch in range(num_epoch):
obs = env.reset()
next_obs = None
reward = 0
total_reward = 0
done = False
traj = Trajectory()
while not done:
if next_obs is not None:
obs = next_obs
obs = torch.Tensor(obs).float()
action = action_decide(net, obs)
next_obs, reward, done, info = env.step(action)
next_obs = torch.tensor(next_obs).float()
traj.add(obs, action, reward, next_obs)
total_reward += reward
# if iteration == 90:
# env.render()
if done:
buff.append(traj)
reward_list.append(total_reward)
A2C(net, optimizer, buff)
env.close()
return reward_list