First, note that we shall use the same MDPs defined in Appendix D.1 as follows

Models like LASSO and ridge regression are extensively used in practice due to their interpretability, ease of use, and strong theoretical guarantees.

In contrast to this, our algorithm works in the composite case with general maximally monotone operatorB(u).

In this section we provide a detailed proof for the main theorem. First we state some facts about the learning rate and the algorithm. This bound contains three parts. The first is an upper bound for the first step when there is no data. The third part is an "average" of the estimated future regret.

Under the assumption thatγGCN 1, i.e., the graph is densely connected and the largest singular value ofweight matrices issmall, we havedM(H(`+1)) γGCN dM(H(`)).

In our method and all the baselines except surrogate-based triage, we use the cross-entropy loss and implement SGD using Adam optimizer [40] with initial learning rate set by cross validation independently foreachmethod andleveloftriageb. Insurrogate-based triage, weusethelossand optimization method used by the authors in their public implementation. Moreover, we use early stopping with the patience parameterep = 10,i.e.,we stop the training process ifno reduction of cross entropy loss is observed on the validation set. This suggests that the humans aremore accurate than thepredictivemodel throughout theentire feature space. This suggests that the humans are less accurate than the predictive model in some regions of the featurespace.