question about paper result

[cycaoyang]

Dear author:
Hello! I am a graduate student in a Chinese university. I am working on a project on multi-agent reinforcement learning. I hope to connect my algorithm to the environment you developed to test the effect. However, when I reproduced the results of your paper, I found something confusing:


These two pictures are the scenes corresponding to L1-33. The first picture is the information recorded by eval(), and the second picture is the information recorded by train(). It can be seen that there is a huge difference in the values ​​between the two. Why is this?

[hsvgbkhgbv]

Hi,

The main reason is that during training the exploration would be considered, e.g., some suboptimal random actions would be conducted. However, during testing the optimal actions are conducted, so the reward would be naturally higher.

Hope this can help you!

[cycaoyang]

Thanks for your reply！

But after reading your code, I have an idea that your control rate indicator is low in training mode because your log_std is fixed when the Gaussian distribution is not adopted by default. When I test the effect of my algorithm, log_std is generated by the network, so the training results and test results are not much different. This may be the reason why your test results are different from the training results?

[Zebei99]

Dear author: Hello! I am a graduate student in a Chinese university. I am working on a project on multi-agent reinforcement learning. I hope to connect my algorithm to the environment you developed to test the effect. However, when I reproduced the results of your paper, I found something confusing: These two pictures are the scenes corresponding to L1-33. The first picture is the information recorded by eval(), and the second picture is the information recorded by train(). It can be seen that there is a huge difference in the values ​​between the two. Why is this?

Sir,
Could you please tell about the visualization package in this graph?