Test results, the acceptance rate is always 1

[lpj12121]

After I trained the model and ran several rounds of tests, I found that the acceptance rate was always 1. The acceptance rate during the validation process during training was about 0.9, which was quite different from the test results. I did not modify the core code, but I could not get results close to those in the paper. Why is this?

[lpj12121]

When I change the parameters of the virtual network, sometimes it will become 0.999. This acceptance rate is too high and it doesn't seem normal.

[GeminiLight]

Hi, could you provide me with more details so I can better understand the problems?

1. Which algorithm do you currently run?
2. How do you set the p_net_setting.yaml and v_sim_setting.yaml?
3. Could you select other baseline algorithms and report their results?

[lpj12121]

Hi, could you provide me with more details so I can better understand the problems?

1. Which algorithm do you currently run?
2. How do you set the p_net_setting.yaml and v_sim_setting.yaml?
3. Could you select other baseline algorithms and report their results?

1.I am verifying HRL-ACRA, and I have trained the upper and lower agents.

2.The p_net_setting.yaml is the default value, which I did not change. The v_sim_setting.yaml is based on the comparative experiments in the paper. I am very curious why the acceptance rate is always 1 or 0.999 during the verification process. During training, there is a verification every ten rounds, and the acceptance rate finally converges to about 0.89.

3.I have trained the A3C-GCN model before, and this is more normal and closer to the data in the paper.

This problem seems to have been mentioned in issue 1, and it looks similar to mine

[GeminiLight]

Hi, the results of my running testing show that the AC rate is not always 1. Could you provide me with more details on any modifications to the original code?

[MerrillLi]

I may know what happens... if you set the decoding strategy to "beam," the AC will have a very high probability of being 1...
Set to greedy can reproduce the results reported in the paper.

[GeminiLight]

Thanks for your feedback. I will check this issue this week.