Definition: Trying to improve the performances of an agent on a target set of which limited data is available by first training on a source set with more data available. Less target data is therefore needed

Data shows that for AVs, single vehicule crash are just 1% (30% for human drivers) of the crashes. It is therefore much more important to include multi agent scenarios.

Risk estimation: depends on severity S, controllability C and exposure E.

S: property damage, injury or death

E: probability of the precrash scenario to happen

C: =0 of the vehicule is never able to prevent crashing =1 when it always prevents it.

Risk R = E * C

During training: interesting to look at the TTC estimated by the algorithm

Scenario library generation: for each logical scenario, a range for each parameter must be tested. New scenarios encountered on the road must be added iif they are estimated to present an unacceptable risk.

It is important to test edge cases where eg senor fusion fails for a given period within the scenario