Horse wrote:jont- wrote:Well, quite. Or at least the average member of the public. My problem with a lot of them is that they are basing ML decisions on data acquired from lots of very average drivers. I want my autonomous car to drive like Andy Morrison.
If it's possible to write that style into rules, then it may well be possible - in time - to achieve that.
Not rules so much as experience. Imagine Andy's thought process on seeing an oncoming emergency vehicle using the "parting the waves" technique (or any other scenario you care to summon up). "I've seen this situation several thousand times. Mostly what happens is x. Other times y or z or a or b or c or d or ... and then there was that one time (oops no twice) when x[23] happened". By the time that sentence is finished, he's also assessed the drivers around him, how they're likely to behave, their age, apparent confidence levels, the vehicles they're driving, the road topology, the state of the verges, presence or not of kerbs, escape routes, has a main plan and two backups, and is putting plan A (ease off with half an eye on the mirror and position far enough left to make space) into operation, ready to adopt plan B or C (or X23) as the situation develops.
The autonomous vehicle will do the physical scan, and assess space, velocity etc, and implement its plan. There will be only one plan, and that won't change until a safety threshold is breached, when it will repeat the process. On implementing its plan, will it take into account its effect on surrounding vehicles? (the guy behind is too close, so I'll apply the brakes rather more gradually than I otherwise might...). And so on ...