21–29 of 29 posts
Earlier quoted context omitted.
True. It's cool. You what else is cool ? an RV. everybody dreams of owning an RV. almost nobody does.
An autonomous RV, now we are talking. :-)
One guess: the routing problem(the vehicle routing problem) is a core computer science problem, for decades. Tons of research have been done about it. I don't expect Google or Uber to have any breakthroughs there, because it's pure luck. But maybe they will, or maybe they'll will ackuire someone. Let's put it down to pure luck. But assuming no brekathrough happens: one company that has tons of experience in that fiel…
For this problem, they don't need to find the optimal routes, just reasonable routes. Individual routes have only four stops at a time. No breakthroughs required.
As for routes only having 4 stops - sure, but that's after you chosen which users will drive in the same car/trip, which is hard in itself.
BTW the dial-a-ride problem is quite similar to the ridesharing problem, and the complexity there is O(number_of_pickup_and_drop_points^2 ) [1]
[1]https://www.itu.dk/people/pagh/CAOS/DARP.pdf - altough it's a bit old, so maybe results have improved since.
Earlier quoted context omitted.
For this problem, they don't need to find the optimal routes, just reasonable routes. Individual routes have only four stops at a time. No breakthroughs required.
Not choosing optimal routes would lead to a more expensive service, which is critical for a commodity. As for routes only having 4 stops - sure, but that's after you chosen which users will drive in the same car/trip, which is hard in itself. BTW the dial-a-ride problem is quite similar to the ridesharing problem, and the complexity there is O(number_of_pickup_and_drop_points^2 ) [1] [1] https://www.itu.dk/people/pag…
Also, this particular problem is much easier than you surmise. The problem statement isn't a set of millions of cars sitting in particular locations with a set of millions of riders with particular sources and destinations and matching them and ordering routes to minimize fuel usage or median rider travel time. The problem is of hundreds of nearby cars traveling existing routes and matching a single new rider with a particular source and destination to one of those cars. There is a secondary problem of where to send empty cars that is more interesting.
Earlier quoted context omitted.
If it will be successful, many will be sold and the price will come down quickly. I don't know why AI should stay expensive.
iPhones are successful, many have been sold, yet the price has not gone down quickly. For autonomous cars, the best-case component price estimates I've seen for solid-state LIDARs, assuming mass production, is $1000. I believe you need four of those, and then you need ultrasound, radar and super-hi-res cameras. And a big honking computer to process all that data. It's probably at least a $10k premium over the normal…
Earlier quoted context omitted.
If it will be successful, many will be sold and the price will come down quickly. I don't know why AI should stay expensive.
iPhones are successful, many have been sold, yet the price has not gone down quickly. For autonomous cars, the best-case component price estimates I've seen for solid-state LIDARs, assuming mass production, is $1000. I believe you need four of those, and then you need ultrasound, radar and super-hi-res cameras. And a big honking computer to process all that data. It's probably at least a $10k premium over the normal…
The real problem is that LIDAR instead of better "stereo camera data processing" is development in hardware rather than brainware (AI).
There is indeed a lot that needs to be developed w.r.t. brainware. However, I'm pretty sure it doesn't need to be recouped through selling cars. Brainware is valuable in many more applications.
1. Accessing the internet through Google vs Viv. 2. Household robots vs cars. 3. Agriculture.
I would be surprised if the planning part is the most difficult part of autonomous cars. Embodied intelligence is really nowhere yet. I've seen Google using particle filters to represent other cars. AI is just not advanced enough to cope with leaves on the road, a pedestrian who wants to cross or not, a broken traffic light, a criminal who want to steal your car, etc. And everything as a service... Really, if we woul…
Earlier quoted context omitted.
Not choosing optimal routes would lead to a more expensive service, which is critical for a commodity. As for routes only having 4 stops - sure, but that's after you chosen which users will drive in the same car/trip, which is hard in itself. BTW the dial-a-ride problem is quite similar to the ridesharing problem, and the complexity there is O(number_of_pickup_and_drop_points^2 ) [1] [1] https://www.itu.dk/people/pag…
If you compute routes that are close enough to optimal, it doesn't matter. If you compute good enough routes quickly, that will likely reduce costs more than computing the best routes slowly. There are enough other areas of optimization that will give a better return on investment that it doesn't necessarily make sense to invest in finding the actual best routes. Also, this particular problem is much easier than you…
But once we start talking about UberPool and via its pretty close for the problem I describe, especially if we're talking about the commute , with a possibility of people preordering rides sometime in advance(or even some prediction abilities about that).
As for the other option, carpooling ,it depends how far will people be willing to drive out of their way for that extra income. But being conservative and I they won't go out of their way , the problem becomes assigning riders to a million bus routes with small capacity, which seems to naturally break this problem into many ,largely independent problems , and may prevent any big exponential complexity.
Don't you agree ?
Earlier quoted context omitted.
If you compute routes that are close enough to optimal, it doesn't matter. If you compute good enough routes quickly, that will likely reduce costs more than computing the best routes slowly. There are enough other areas of optimization that will give a better return on investment that it doesn't necessarily make sense to invest in finding the actual best routes. Also, this particular problem is much easier than you…
I agree. But that's true for current Uber . But once we start talking about UberPool and via its pretty close for the problem I describe, especially if we're talking about the commute , with a possibility of people preordering rides sometime in advance(or even some prediction abilities about that). As for the other option, carpooling ,it depends how far will people be willing to drive out of their way for that extra…
Earlier quoted context omitted.
If you compute routes that are close enough to optimal, it doesn't matter. If you compute good enough routes quickly, that will likely reduce costs more than computing the best routes slowly. There are enough other areas of optimization that will give a better return on investment that it doesn't necessarily make sense to invest in finding the actual best routes. Also, this particular problem is much easier than you…
I agree. But that's true for current Uber . But once we start talking about UberPool and via its pretty close for the problem I describe, especially if we're talking about the commute , with a possibility of people preordering rides sometime in advance(or even some prediction abilities about that). As for the other option, carpooling ,it depends how far will people be willing to drive out of their way for that extra…
That particular users have commutes at around the same time every weekday, plus or minus half an hour depending on their first meeting and when they finish their work or close their shift or socialize with coworkers, will be better handled by the system placing empty cars appropriately.