Wealso provide improvedregret bounds when theothers' maximum bid exhibits the further structure of sparsity.

This lemma is a trivial consequence of the definition of Meijer G-functions. The only nontrivial step in the abovereasoning is going from the second to the third line. To speed up the process, the experiments are done by using a restriction ofGH excluding the inverse trigonometric functions as well as some Bessel functions. Also note that, as suggested by LIME,X8, X9 also have an important weight in this polynomial. Wefinishonalast remark on the benefits offered by our projection pursuit approach. Wesee that both the symbolic model and its local approximation take a very concise form when we consider the new variables zk, k =1,...,K.

Machine Learning has proved its ability to produce accurate models - but the deployment of these models outside the machine learning community has been hindered by the difficulties of interpreting these models.

For our synthetic datasets, we conduct two additional rounds of instruction evolution to increase complexity. As shown in Figure 3, the scores of instructions across the three datasets consistently increase.