How were the reward hyperparameters obtained?
ManifoldFR opened this issue · 2 comments
ManifoldFR commented
Hi,
I read the paper and looked through the code. I am wondering how the parameters (lengthscales, for instance) for the reward kernels and the reward tree were obtained. Was it some rule of thumb?
I guess random search or using optuna is a possibility, but that would be expensive, right?
Thanks.
Jungdam commented
Hi,
In case of some reference values are available (e.g., similar terms in Peng 2018 or Park 2019), I first adopted the values in those references then modified them accordingly. Otherwise, I choose parameters experimentally from scratch, where my rule of thumb is to make all the terms be equal as possible then modify the values according to the achieved results.
ManifoldFR commented
Thanks !