environment？

[xss1006]

Hello, thank you for your work!
I would like to ask, when normalizing the state, what is the basis for setting the values? For example:
delta_yaw_t = np.array(self.state_info['delta_yaw_t']).reshape((1, )) / 2.0
dyaw_dt_t = np.array(self.state_info['dyaw_dt_t']).reshape((1, )) / 5.0
lateral_dist_t = self.state_info['lateral_dist_t'].reshape((1, )) * 10.0
action_last = self.state_info['action_t_1'] * 10.0
future_angles = self.state_info['angles_t'] / 2.0

What do the values you divide or multiply by represent? Your answer would be of great help to me. I look forward to your response.

[ShuaibinLi]

No specific physical meaning, the purpose is to normalize the state.

[ShuaibinLi]

Similar to action normalization in continuous envs

[xss1006]

Similar to action normalization in continuous envs
Thank you for your reply，If I set it to different values, will it have a significant impact?

[ShuaibinLi]

Similar to action normalization in continuous envs
Thank you for your reply，If I set it to different values, will it have a significant impact?

maybe, worth trying! Hope for better results.

[ShuaibinLi]

https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/blob/master/a2c_ppo_acktr/envs.py some hints?

[xss1006]

https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/blob/master/a2c_ppo_acktr/envs.py some hints?

Thank you for your suggestion. I'll take a look at it later. I noticed that in your code, the states are not normalized to a range of [0, 1] or [-1, 1], but instead, they are divided by a value to make them fall within a different range, right?

[ShuaibinLi]

https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/blob/master/a2c_ppo_acktr/envs.py some hints?

Thank you for your suggestion. I'll take a look at it later. I noticed that in your code, the states are not normalized to a range of [0, 1] or [-1, 1], but instead, they are divided by a value to make them fall within a different range, right?

Due to their different meaning, yes. Actually, I didnot do many experiments here

[xss1006]

https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/blob/master/a2c_ppo_acktr/envs.py some hints?

Thank you for your suggestion. I'll take a look at it later. I noticed that in your code, the states are not normalized to a range of [0, 1] or [-1, 1], but instead, they are divided by a value to make them fall within a different range, right?

Due to their different meaning, yes. Actually, I didnot do many experiments here

Sure, may I ask, if the states are simply converted from one range to another, what is the significance of normalize state？

[ShuaibinLi]

https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/blob/master/a2c_ppo_acktr/envs.py some hints?

Thank you for your suggestion. I'll take a look at it later. I noticed that in your code, the states are not normalized to a range of [0, 1] or [-1, 1], but instead, they are divided by a value to make them fall within a different range, right?

Due to their different meaning, yes. Actually, I didnot do many experiments here

Sure, may I ask, if the states are simply converted from one range to another, what is the significance of norm

Great question!! To improve the efficiency and stability of learning, states & actions are suggested to be normalized， especially in some dynamic envs (eg mujoco). While to [0,1] or [-1,1]，actions maybe a good choice, I think it is best not to normalize the state too much for better generalization (need experiments).

[xss1006]

I would like to know more about why you divide by these specific values when processing different states. What is the reasoning behind choosing these particular numbers?

https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/blob/master/a2c_ppo_acktr/envs.py some hints?

Thank you for your suggestion. I'll take a look at it later. I noticed that in your code, the states are not normalized to a range of [0, 1] or [-1, 1], but instead, they are divided by a value to make them fall within a different range, right?

Due to their different meaning, yes. Actually, I didnot do many experiments here

Sure, may I ask, if the states are simply converted from one range to another, what is the significance of norm

Great question!! To improve the efficiency and stability of learning, states & actions are suggested to be normalized， especially in some dynamic envs (eg mujoco). While to [0,1] or [-1,1]，actions maybe a good choice, I think it is best not to normalize the state too much for better generalization (need experiments).

Thank you very much for your reply; it has been very helpful. When you normalize the state, is it based on experience, like in the case of this: delta_yaw_t = np.array(self.state_info['delta_yaw_t']).reshape((1, )) / 2.0? Can I more about why you divide by these specific values when processing different states. What is the reasoning behind choosing these particular numbers?