some small experiments on learning to sample the swiss roll distribution while optmimizing the beta schedule