Problem about over fit

[thevelvetunderground]

Dear Pro. Lu
I use PINN to solve a Laplace's equation：

I set 30000 iterations for Adam and the L-BFGS. Then I continue training with the RAR method. there are 9600 points in the study area.
And I sampled 6000 points. But it always has over fit problem. How can I fix this problem?

[tsarikahin]

I believe this is a bug in DeepXDE. Let me explain. The training data includes both boundary and collocation points. However, test data has only collocation points collected inside the domain (function tes_points of pde.py). If bc error is large, then you will end up with a larger training error than the test error. What @lululxvi says about this?

[lululxvi]

1. I don't see where is overfitting.
2. It is not a bug. Test data also contains the same boundary points in training.