Benchmarking vision-based driving policies is challenging. On one hand, open-loop evaluation with real data is easy, but these results do not reflect closed-loop performance.

To this end, we follow Ainsworth et al.

Toourknowledge, ADASPIDER isthefirstparameterfree non-convex variance-reduction method in the sense that it does not require the knowledge of problem-dependent parameters, such as smoothness constant L,targetaccuracyϵoranybound ongradient norms.

This paper will focus on the case of graph classification, but with minor modifications, our framework can beextended toother tasks ofinterest.