Tricks for fine-tuning the hyperparameters

[liushunyu]

Hello! I am trying to implement other reinforcement learning method to deal with the l2rpn problem, but I find that my result cannot match the performance with the DQN implementation in l2rpn-baselines.

Even I use the other library to implement the DQN, when I change the hyperparameters a little, I cannot get the reasonable result.

So I want to know how you fine-tune the hyperparameters of DQN in the l2rpn problem. Is there any tricks?

[Tezirg]

Hello,

Fine tuning NN hyperparameters is a whole research area and there is no definitive answer. For the DQN, it has been done by trial and error, alongside rigorous data tracking in both spreadsheets and tensorboard. A good starting point is also to use the hyperparameters of the reference literature.

Finally, I would consider the DQN more like an example of how to get started with L2RPN+NN approach, not like a really good performing baseline. This is because the L2RPN challenge is more about solving the increasing complexity as the grid size grows from 10, to 100, 1000, and 10000 nodes; which a DQN cannot handle with its 1:1 mapping of output to actions.

Best