On Tue, 2019-05-21 at 11:33 +1000, Bernard Robertson-Dunn wrote: > In my experience as an automation engineer, when you automate > something you have to be able to deal with all the exceptions that a > human can manage instinctively. Humans can usually tell that > something is unexpected and has a strategy to deal with it - stop, > swerve etc. This is not 100% fail safe but it has a reasonable track > record. Machines need to be proactively programmed. This is hard.
Humans don't manage much instinctively, or young drivers would be a whole lot better than they are. The idea that human have a "reasonable track record" is laughable - they kill and maim tens of thousands of each other each year. If machines had the same track record, we wouldn't let them any where near the roads. In fact they have a much better track record, and we still don't let them drive. People learn, and it is a whole lot slower than if we could just upgrade them. Often, what would be a learning experience for a machine is a dying experience for a human. Humans are really, REALLY bad at learning from other people's experiences. > Autodriving/full self-driving cars need to be tested against > exceptions, not the norm. AFAIK that has never happened and may only > happen over time in use, not the lab. I don't think the Tesla engineers are stupid. Being safe is an existential requirement for Tesla. I'm pretty sure their lab testing is as comprehensive as they can make it - including testing against exceptions. Bear in mind that given the billions of kilometres travelled by cars every year, most "exceptions" happen thousands of times per day. Plus Tesla has been running their cars on real roads with real drivers for years now. Regards, K. -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Karl Auer ([email protected]) http://www.biplane.com.au/kauer http://twitter.com/kauer389 GPG fingerprint: 8D08 9CAA 649A AFEF E862 062A 2E97 42D4 A2A0 616D Old fingerprint: A0CD 28F0 10BE FC21 C57C 67C1 19A6 83A4 9B0B 1D75 _______________________________________________ Link mailing list [email protected] http://mailman.anu.edu.au/mailman/listinfo/link
