sandyhsia / dqn-gym

Experiments on DQN

DQN-gym

Experiments on DQN

Environment request:

Tensorflow 0.10.0

(Other version might need you change some function usages)

Playing with your AI

1. To train:

python ./myRL.py

2. To display:

python ./myRL-display.py

3. Hyper-params in myRL.py or myRL-display.py:

CHECKPOINT_DIR = './checkpoint

where you can save to, or load from check point.

save_request = 1

means you're sure your environment is ok, thus you want to save your model after a period of training.

restore_request = 1

means you want to train from "this" check point file (clarified in './checkpoint/checkpoint' file) Usaully, the file contain the lastest checkpoint file name. However, in another way around, if you change the the file name here, then, it would start from the place you want.

4. Where to change:

1. (Strongly recommend...) Every hyper-params in car-DQN.py and Q-network structure here.

2. reward_method() in Virtual-Env.py

3. In car-DQN.py, I am still not sure how to manipulate the replay_buffer(). And I am trying to add so-called memory and attention to agent.

Idea from here: https://zhuanlan.zhihu.com/p/21320865?refer=intelligentunit