wujingda/human-in-the-loop-deep-reinforcement-learning

(Engineering) Toward human-in-the-loop AI: Enhancing deep reinforcement learning via real-time human guidance for autonomous driving

Human-in-the-loop Deep Reinforcement Learning (Hug-DRL)

This repo is the implementation of the paper "Toward human-in-the-loop AI: Enhancing deep reinforcement learning via real-time human guidance for autonomous driving".

Toward human-in-the-loop AI: Enhancing deep reinforcement learning via real-time human guidance for autonomous driving

Jingda Wu, Zhiyu Huang, Zhongxu Hu, Chen Lv

Getting started

1. Install the CARLA simulator (0.9.7), with referring to https://carla.readthedocs.io/en/latest/start_quickstart/#a-debian-carla-installation

2. Install the dependent package

pip install -r requirements.txt

3. Run the training procedure

python train_offline.py

Training performance of different algorithms

Reference

If you find this repo to be useful in your research, please consider citing our work

@article{WU2022,
title = {Toward human-in-the-loop AI: Enhancing deep reinforcement learning via real-time human guidance for autonomous driving},
journal = {Engineering},
year = {2022},
issn = {2095-8099},
doi = {https://doi.org/10.1016/j.eng.2022.05.017},
author = {Jingda Wu and Zhiyu Huang and Zhongxu Hu and Chen Lv},
}

License

This repo is released under GNU GPLv3 License.

human-in-the-loop-deep-reinforcement-learning's Issues

论文程序请教

吴同学，

您好！

您发表的论文《Toward human-in-the-loop AI: Enhancing deep reinforcement learning via real-time human guidance for autonomous driving》十分精彩。也非常感谢您无私的分享了论文的相关程序，我在学习您的算法时，可能发现了程序的一个小问题。主要是当算法训练的一个episode结束后，需要更新replay memory buffer时，errors在脚本中没有被定义。我附上了相应的报错信息。

File "...\TD3_based_DRL\TD3.py", line 115, in learn
self.memory.batch_update(tree_idx, abs(errors.detach().cpu().numpy()) )
NameError: name 'errors' is not defined

不知您是否有空看一看程序这个部分？非常感谢您的帮助。

a bug about critic update

Hello,

I find your work is really helpful and I really appreciate it, however I found a bug at critic update stage which affect the final performance.

It is in TD3HUG.py at L80-L81

noise1 = (torch.randn_like(ba) * self.policy_noise).clamp(0, 1)
a_ = (self.actor_target(bs_).detach() + noise1).clamp(0, 1)

I think the first parameter in clamp should be the lower limit rather than 0, and I think noise1 should use the NOISE_CLIP hyperparameter.

I change these two lines into

noise1 = (torch.randn_like(ba) * self.policy_noise).clamp(-self.noise_clip, self.noise_clip) # self.noise_clip refer to the NOISE_CLIP hyperparameter
a_ = (self.actor_target(bs_).detach() + noise1).clamp(-1, 1)

There might be also bugs about clamp elsewhere, but I didn't check.

Would be very appreciated if you look into this.

question about actor loss

Hi,
In the paper, actor loss is

but the code that calculate actor loss for human intervention steps didn't consider the first term (see https://github.com/wujingda/Human-in-the-loop-Deep-Reinforcement-Learning/blob/main/TD3_based_DRL/TD3HUG.py#L148)
Also, the human intervention weight in actor loss in the code has a soft update coefficient which isn't included in the paper, and I don't understand what this coefficient is for (see https://github.com/wujingda/Human-in-the-loop-Deep-Reinforcement-Learning/blob/main/TD3_based_DRL/TD3HUG.py#L144).
Are these bugs in the code or some tricks that I misunderstand?
Looking forward for your help.