marginnote3app://note/1A5236FF-C4A7-4FA4-920E-456FC2E587FF michigan has AV test facility intelligent transpo system - safety validation of AV - mcity test facility is first purpose built proving ground for connected. automated vehicles - winter is coming for av which forecasted 10million (drop of valuation: remo, auror, zooks) - fatality is 1/10^6 - weather, road infrastructure, road users, maneuvers, - av must give way to emergency vehicle even it has the right of way - 10^-4 large imbalance ratio - expectation wouldn't change - converge to same accident rate, real vehicle (virtual vehicles are controlled ) - quantum signal ai (from ford) I just differentiate optimizing for the variance but not the mean; why should the agent derivation of reward is coming from importance sampling (actual function is from neural network) - p.31: what, when, how: 15% of adversearial behavior; can we just cut and paste adverserial behavior?