Minimax strategy and early_stopping

[Evey-AI]

Hi, I have two issues with Minimax strategy and early stopping : 1, loss1 is to maximize series_association and loss2 is to minimize prior_association. In the original paper, it was minimize, then maximize. Why did it become maximize, then minimize in the code?

2. Why score = -loss when stopping early？

[zixuan-go]

Come uninvited : ). The following is my own understanding, for reference only, correctness not guaranteed： *Please note that the order in the code is Max-Min, not Min-Max as in the paper.

The reason why Min-Max cannot be used is that the random initialization of series association (SA) has no meaning and may be very far away from the target point, which makes it difficult to continue training. In extreme cases, the model can no longer use more distant context for modeling.

The initialization of prior association (PA) is affected by the unimodal statistical properties of its Gaussian kernel function, so there is a natural neighborhood meaning and it can be directly used as the Max-training label.

That's why we use Max-Min rather Min-Max. Hope for helping.