Pokemon Challenge using NN

#1

I tried to make a code for the pokemon challenge by CB using NN, 2 hidden layers(100, 50) and learning rate=0.0002 (tried varying from 0.01 to 0.0002).

I’ve changed the number of iterations all the way from 100 to 1000, but the loss stops reducing at about 0.17 or 0.18.

Why is the loss not reducing any further?
How do I reduce the loss further?

#2

hey @Aayushhh ,
The reason for loss to not be lowering more is that either the model has reached it’s convergence or is stuck at any local minima , and as your are using very low learning rate , then it is also not able to come out of that local minima point.

What you can do to improve is,

  1. although we know that we are currently stuck at local minima , but we can try using more lower learning rate with increased number of iterations , this can help the model to come out of local minima.
  2. Try getting more data or improving your model by adding more layers or something else modifications.
  3. Try model hyper parameter tuning. ( means changing the number of nodes in each hidden layer and taking the best combinations that performs best. )
  4. Changing model optimizer

I hope these points may help you .
Thank You and happy Learning :slightly_smiling_face:.