Some questions about the experimental.

[FelixHuiweiLin]

1.How to calculate the global optimum p ∗ in the relative suboptimality plots and regret plots?
2.During your model training, do you use training data mini-batch or one by one？

[mmahesh]

1. You can run some known algorithm such as Newton method or LBFGS or Gradient Descent for a long time then obtain p*, if the problem is convex. You can then use that p* in relative suboptimality plots or regret plots. For nonconvex problems, as a heuristic you can use the least function value attained by any of the algorithms (if you use sufficiently large number of iterations).
2. We used mini-batching in general, however, for regret plots we use one by one.

[FelixHuiweiLin]

Thanks!