Open cosmicBboy opened 3 years ago
To implement this, need to generalize the hyperparameter head to support both multiclass classification over a discrete and finite hyperparameter space, but also a continuous hyperparameter space.
For the continuous case, the heads should be mu
and sigma
that produce an inference of the parameters that govern the shape of a gaussian. On policy generation, these two values should be used to parameterize a normal distribution that we sample from to produce a scalar of the continuous action (e.g. the l2 regularization parameter value).
Support continuous action space for selecting real hyperparameters within the bounds specified by algorithm space config: