I'm currently using Weights & Biases (W&B) to track and optimize hyperparameters for my algorithm. I have two hyperparameters, radius and k_neighbors, which are correlated in such a way that a smaller radius corresponds to a larger number of neighbors (higher k_neighbors), and vice versa.
I would like to perform a parameter search using the random method in W&B's sweep feature, but I want to ensure that the generated parameter combinations respect the desired relationship between radius and k_neighbors. In other words, I want to have a higher probability of selecting parameter combinations where a larger radius is paired with a smaller k_neighbors, and a smaller radius is paired with a larger k_neighbors.
Is there a way to achieve this in W&B? Specifically, I'm looking for guidance on how to define and incorporate a joint distribution or copula function for these two hyperparameters in the context of W&B's sweep. My goal is to sample parameter combinations during the sweep that adhere to the desired correlation between radius and k_neighbors.
I have already explored the W&B documentation, but I couldn't find specific guidance on how to implement this joint distribution or copula approach in the context of a sweep. Any suggestions, code examples, or pointers to relevant resources would be greatly appreciated.
Thank you in advance for your help!
