Q_target = reward + self.gamma * Q_target
self.policy_optimizer.zero_grad()
loss = F.mse_loss(Q_current.float(), Q_target.float())
loss.backward()
self.policy_optimizer.step()
if __name__ == '__main__':
env = gym.make('CartPole-v0')
state_space = env.observation_space.shape[0]
action_space = env.action_space.n # num_action env.action_space.shape[0]
max_episode_steps = env._max_episode_steps
policy = DQN(state_space, action_space)
episode_reward = 0
episode_timesteps = 0
capacity = 100000
use_replay_buffer = 1000
batch_size = 32
replay_buffer = ReplayBuffer(state_space, action_space, capacity)
state, done = env.reset(), False
for t in range(int(10000)):
episode_timesteps += 1
action = policy.select_action(state, t)
#action = select_action(state)
next_state, reward, done, _ = env.step(action)
not_done = float(done) if episode_timesteps < max_episode_steps else 0
episode_reward += reward
replay_buffer.store(state, action, next_state, reward, not_done)
state = next_state
if replay_buffer.buffered(64):
policy.train(replay_buffer)
