The difference in performance between the reported results in the paper and the actual implementation of the GA3C-CADRL approach in this repo

[zhangqq21]

Hello, I encountered an issue while running the code from this repository. When testing cases with 8 agents navigating randomly, I found that the success rate of the CADRL method is 83%, which matches the results in the paper. However, the success rate of GA3C-CADRL is only 39%, which does not align with the paper, where the reported success rate is over 90%. Could you please explain the reason for this? I am using the neural network model parameters provided in the repository.