4 ms·

#OldNews, this article is from 2016. > What sets ALPHA apart is the genetic fuzzy tree decision system that can calculate an opponent’s movements or strategies

#OldNews, this article is from 2016. > What sets ALPHA apart is the genetic fuzzy tree decision system that can calculate an opponent’s movements or strategies 250 times faster than you can blink. That's just not fair. One more comment, can this thing (ALPHA) land an Airbus A320 on the Hudson?

tempestn 9y ago
Not yet, but I can't imagine a future where AIs don't eventually control takeoff, landing, and emergency situations.
phire 9y ago
Water landings are a documented procedure for planes, there is a checklist (which Sully had practiced in simulators). Assuming an AI like this was given 100% control over a commercial plane, it would have been programmed with such a checklist. It would have also been tested on this ability in a simulator (just like regular pilots). The AI would have easily worked out in an emergency that it couldn't reach the runway, or any alternative runway and not even bothered trying. The only question is, would the AI have decided that the Hudson was a valid surface for a water landing attempt? And that really depends on the database programmed into the plane. It can't really see the river, so the database would need to have the river marked as water, at the correct elevation. It also can't really see the bridges (maybe if it had LIDAR), so the bridges would have to be included in the database too, so it could pick a good stretch to try and land it. Oh and boats too... either it would have to dynamically spot them, or get lucky. But if the database correctly reflected reality, then it would see the opportunity, try to execute it and potentially succeed.
- autokad 9y ago
  AIs are still struggling with 'is that white space a tractor trailer or sky? eh just drive through it' with cars. they are immensely far off from aerial combat and emergency maneuvers in aircraft. and the whole point of emergencies is that they are a bit unexpected and the situations tend to be rather specific. I'm not saying it wont eventually happen, but no: they unequivocally cannot fly our aircraft in takeoff/landings, emergency situations, etc.
- astrodust 9y ago
  Don't forget that humans can still set the strategy and the AI can figure out the details of implementation. If you say land on an iceberg, it might be able to compute a viable solution.
- davidwihl 9y ago
  Most aviation accidents are due to pilot error. I would rather have a debugged AI reduce overall accident rates than the once in a generation Sully heroics. Most commercial flights are already on the autopilot from 50' after takeoff until taxiing after landing. Already in CAT III landings or above FL280, the autopilot, not the human pilot, must be flying. Expect more of this in additional phases of flight.
astrodust 9y ago
It would depend on the software. Pilots can only rely on their own personal experience, plus anything they might have been taught, overheard, or read about somewhere. An AI pilot can learn from every mistake made by every AI pilot, can absorb everything important from their collective experience. It's not a fair fight. If the designer set about testing the AI's ability to land on water, eventually it'd find a way of doing it, and then that could become part of the corpus of experience. AI pilots stand to theoretically be better than any pilot ever. If there's ever a better human pilot, their experience can be documented, examined, and imported into the core AI.

tempestn 9y ago

Not yet, but I can't imagine a future where AIs don't eventually control takeoff, landing, and emergency situations.

phire 9y ago

Water landings are a documented procedure for planes, there is a checklist (which Sully had practiced in simulators). Assuming an AI like this was given 100% control over a commercial plane, it would have been programmed with such a checklist. It would have also been tested on this ability in a simulator (just like regular pilots). The AI would have easily worked out in an emergency that it couldn't reach the runway, or any alternative runway and not even bothered trying. The only question is, would the AI have decided that the Hudson was a valid surface for a water landing attempt? And that really depends on the database programmed into the plane. It can't really see the river, so the database would need to have the river marked as water, at the correct elevation. It also can't really see the bridges (maybe if it had LIDAR), so the bridges would have to be included in the database too, so it could pick a good stretch to try and land it. Oh and boats too... either it would have to dynamically spot them, or get lucky. But if the database correctly reflected reality, then it would see the opportunity, try to execute it and potentially succeed.

autokad 9y ago
AIs are still struggling with 'is that white space a tractor trailer or sky? eh just drive through it' with cars. they are immensely far off from aerial combat and emergency maneuvers in aircraft. and the whole point of emergencies is that they are a bit unexpected and the situations tend to be rather specific. I'm not saying it wont eventually happen, but no: they unequivocally cannot fly our aircraft in takeoff/landings, emergency situations, etc.
astrodust 9y ago
Don't forget that humans can still set the strategy and the AI can figure out the details of implementation. If you say land on an iceberg, it might be able to compute a viable solution.
davidwihl 9y ago
Most aviation accidents are due to pilot error. I would rather have a debugged AI reduce overall accident rates than the once in a generation Sully heroics. Most commercial flights are already on the autopilot from 50' after takeoff until taxiing after landing. Already in CAT III landings or above FL280, the autopilot, not the human pilot, must be flying. Expect more of this in additional phases of flight.

astrodust 9y ago

It would depend on the software. Pilots can only rely on their own personal experience, plus anything they might have been taught, overheard, or read about somewhere. An AI pilot can learn from every mistake made by every AI pilot, can absorb everything important from their collective experience. It's not a fair fight. If the designer set about testing the AI's ability to land on water, eventually it'd find a way of doing it, and then that could become part of the corpus of experience. AI pilots stand to theoretically be better than any pilot ever. If there's ever a better human pilot, their experience can be documented, examined, and imported into the core AI.