Unmanned Mission to Alpha Centauri, A study of an unmanned mission to the Alpha Centauri system Options

[Greg Hullender]

There's a good bit of serious information on the JPL AI group's page:

http://ai.jpl.nasa.gov/public/

In particular, there is/was a satellite, EO-1 (Earth Observing One), which is using some of their planning algorithms to select science targets. In this paper, they claim a 100-fold increase in science return, measured as science events per megabyte:

http://ai.jpl.nasa.gov/public/papers/sherw...06_autonomy.pdf

From reading this, though, it required at least weekly human intervention; given a broad plan, it could work out the details, but it couldn't extrapolate from one week's results to construct a good plan for the next one. (Assuming I'm reading this correctly; planning algorithms matter a lot in my own research, but this paper isn't actually about the algorithms -- it's much higher level.)

Things like the Mars Rovers use custom-built planning software, but the JPL AI team is actually trying to build a platform that future missions can work from. That'll be really great if it works. Best of all, it'll involve papers I can even justify reading! :-)

--Greg

[ilbasso]

One other comment about AI software updates - in flight software updating works well. However, experiments in advancing AI are demonstrating that the basic architecture of the computer system (e.g., massively parallel processors) could well need to be modified to more closely mimic the functioning of an intelligent nervous system. At least as far as I understand, you can't do that kind of redesign once the probe is launched.

[JRehling]

[...]

[imipak]

Full-featured Artificial Intelligence: Hard, with a capital H.
Artificial intelligent-enough-to-cope-with-this-well-bounded-problem-domain: relatively easy.

It's still very hard in absolute terms, of course, but it's not Hard in the CS (or AI or philosophy of consciousness) sense. I agree with JRehling; we could conceivably build computers to handle the problem with today's technology, given sufficient budget.

Going back to basic mission architecture -- one way of overcoming the problem of slowing down enough to enter solar orbit might be not to do so. Instead the relatively heavy, fast-moving mothership could jettison several very light planetary probes. These could then brake with $[sufficiently_advanced_technology] of choice, without having to slow lots of superfluous superstructure designed for the multi-decade deep space portion of the mission. Perhaps that braking wouldn't be enough to reach planetary orbit, but stretching a (say) 10 day encounter to 100 or 1000 days would be good. Maybe these small vehicles could use the powerful transmitter of the (still relatively close) mothership to relay comms back to earth... though that suggests a DSN-scale antennae on the mothership to receive from the probes. :?

[stevesliva]

Artificial intelligent-enough-to-cope-with-this-well-bounded-problem-domain: relatively easy.

You are looking for the word "autonomy." Or autonomous.

[Greg Hullender]

Just to put a word to it, we say a problem is "AI-hard" to mean the computer would need to be able to think at least as well as a human being in order to solve it. At present, we are nowhere close to building a true artificial intelligence, so if a problem is AI-hard, that means we cannot solve it even in theory.

I agree with JR that this problem isn't "AI-hard". As I posted earlier, the EO-1 satellite seems to have demonstrated that JPL already has software to find scientifically-interesting targets. In fact, odds are good that that software will be very important on future missions to places Neptune, since it lets you make much better use of your precious communications bandwidth.

My own research, on "Practical Dialogue Systems" is aimed at finding a useful, tractable subset of the AI-hard "General Dialogue Problem," so this is the kind of stuff I think about/work on every day, albeit not in a space context.

--Greg

[Phil Stooke]

"the EO-1 satellite seems to have demonstrated that JPL already has software to find scientifically-interesting targets."

Admittedly, I know nothing about this, but I would imagine that the EO-1 software referred to is based on a list of pre-defined things that might be interesting, which could be hard to adapt to a reconnaissance of a new planetary system.

I imagine the procedure would have to be something like this:

1. Repeated surveys during approach to identify planets, with more added to the list as the spacecraft gets nearer (obviously easy to do, Voyager-like).

2. Track objects from survey to survey and compute orbits (should be relatively easy).

3. Repeat 1 and 2 in higher resolution sequences to characterize satellite systems (less time to act but should be relatively easy).

4. Attempt to characterize the objects: size, albedo, temperature, which have rings, which have atmospheres (presumably this can be done, but how soon before the flyby?)

5. Prioritize the highest resolution observations (that looks like the hardest part to me, without human intervention).

But I repeat, I know nothing!

Phil

[Greg Hullender]

Admittedly, I know nothing about this, but I would imagine that the EO-1 software referred to is based on a list of pre-defined things that might be interesting, which could be hard to adapt to a reconnaissance of a new planetary system.
.
.
.

5. Prioritize the highest resolution observations (that looks like the hardest part to me, without human intervention).

This would probably be a more interesting discussion in the context of something like the proposed Argo Neptune flyby -- or even just the Jupiter Trojan flyby portion of that mission. Heck, even New Horizons might be able to use something similar. (We should ask Alan if they have considered anything like the EO-1 software.)

--Greg

[djellison]

NH doesn't need it. They know roughly what they're going to see, and they need to schedule each and every tiny second of that flyby to get the data they need by hand.

[Greg Hullender]

NH doesn't need it. They know roughly what they're going to see, and they need to schedule each and every tiny second of that flyby to get the data they need by hand.

Perhaps. On the other hand, suppose, for example, some unusual feature were seen during approach. The software could (in theory) make sure it got shots of it during closest approach. Or if it spotted another satellite, it could try to get a shot at close approach -- or at least get shots at different phase angles. Yeah, that's weak. :-)

I'll admit the lesson of EO-1 seems to be more for orbiters -- since that's what it was. The claim that it increased the quantity of scientifically-interesting events by a factor of 100 really caught my eye. A Neptune orbiter might have lots of trouble getting enough bandwidth to send back a steady stream of video, but if the software could intelligently pick and choose, that might allow the use of a smaller antenna -- or a higher-resolution camera.

In the case of a flyby, though, perhaps it's just not that hard to simply store everything and then send the results back over time. I suppose, then, what makes the proposed Alpha Centauri flyby unique is the assumption that the designers would have no idea at all what to expect. That the probe would have to choose all of its targets by itself.

--Greg

[jsheff]

For me, Project Longshot would carry a lot more scientific credibility if they hadn't made the amateurish error of designating as Beta Centauri the star that should properly be referred to as Alpha Centauri B. The destination star system is composed of Alpha Centauri A, Alpha Centauri B, and Alpha Centauri C (or Proxima Centauri), all revolving around their common center of gravity. Beta Centauri is a different star system entirely - hundreds of light-years beyond Alpha Centauri and unrelated to it aside from being by chance located in roughly the same part of the sky as seen from Earth, and hence designated by us - arbitrarily - to be in the same constellation. That's an elementary error that seems to carry right through the Project Longshot report.

I hope they're not as sloppy in designing the fusion engines.

- John Sheff
Cambridge, MA

[Hungry4info]

For me, Project Longshot would carry a lot more scientific credibility if they hadn't made the amateurish error of designating as Beta Centauri the star that should properly be referred to as Alpha Centauri B.

Well... I had the same confusion when I was younger...

... like the age of six or seven

[Stephen]

Just to put a word to it, we say a problem is "AI-hard" to mean the computer would need to be able to think at least as well as a human being in order to solve it. At present, we are nowhere close to building a true artificial intelligence, so if a problem is AI-hard, that means we cannot solve it even in theory.

Since this discussion still seems to be sputtering along I'll add another 2 cents.

Claiming a problem as being unsolvable "even in theory" is to claim that no solution for that problem exists. An obvious example would be trying to build a space probe which could ATTAIN light speed.

On the other hand, having a problem which was merely very hard to solve--and the statement that researchers were "nowhere close to building a true artificial intelligence" would seem to be inferring that--is a different beastie altogether. An example would be trying to build a space probe which could reach very high rates of c. Such a probe would doubtless be very difficult to build, and would probably require technology, materials, techniques, etc not available or even feasible today. In the end it may even turn out to be so difficult (or at least so costly) that it may not be worth the effort of trying to build such a probe. But being hugely difficult and impractical that is NOT the same as being impossible.

Given that context, I would argue your statement was self-contradictory. Being "nowhere close to building a true artificial intelligence" is not the same as being unable to build one "even in theory". Has anyone ever demonstrated that being able to build a "true artificial intelligence" was impossible "even in theory"?

I agree with JR that this problem isn't "AI-hard". As I posted earlier, the EO-1 satellite seems to have demonstrated that JPL already has software to find scientifically-interesting targets.

Glad to hear it. I take it then that it was Opportunity rather than Steve Squyres and his team who made the recent decision to head off for Endeavour crater.

======
Stephen

[Stephen]

I don't see any requirements for this task that inherently require a massively parallel architecture, or any particular architecture whatsoever.

Well, I guess that depends on whether artificial intelligence researchers are trying to solve the problems they face by reinventing the wheel or by finding out first how Mother Nature has solved those same problems then using her solution(s) as at least the starting point for their own.

After all, it only took Mother Nature several billion years (using her trial and error methods of natural selection) to figure out how to go about it even if we're only talking about intelligence of the level of a fish or a trilobite. I imagine modern researchers will be hoping to find their own solutions a tad sooner.

======
Stephen

[JRehling]

[...]