[X] My Assistant

[YesRushGen]

Hi all, I am a frequent lurker here at UMSF and rarely post. I came across the following link today and I found it a fascinating read. I thought I would share for any interested persons who had not read this before.

http://ntrs.nasa.gov/archive/nasa/casi.ntr..._1989007533.pdf

Cheers!

Kelly

[Juramike]

Project Longshot (from link above) proposed using existing technology to get to the Centauri system and would take about 100 years.
(faster loading summary here: http://en.wikipedia.org/wiki/Project_Longshot )

Project Daedalus, proposed about 10 years earlier, would've used not-yet-invented technology to get to the Centauri system and it would have had a flight time of 50 years. Here's a quick summary link: http://en.wikipedia.org/wiki/Project_Daedalus

(Interesting that existing technology would only double the proposed flight time to the nearest star. The key quote from Wiki was that "although some technological development would still be required" - probably for the inertial confinement fusion thruster.)

I'm guessing the selection and funding process would also take at least 100 years....

[Stephen]

Project Longshot (from link above) proposed using existing technology to get to the Centauri system and would take about 100 years.

<snip>

Project Daedalus, proposed about 10 years earlier, would've used not-yet-invented technology to get to the Centauri system and it would have had a flight time of 50 years.

For the record, Project Daedalus would have targeted Barnard's Star, a red dwarf, not the Alpha Centauri system (G type + K type + a (distant) red dwarf).

That aside it could also be argued that both projects would be relying (in one way or another) on "not-yet-invented technology".

Due to the great distance at which the probe will operate, positive control from earth will be impossible due to the great time delays involved. This fact necessitates that the probe be able to think for itself. In order to accomplish this, advances will be required in two related but separate fields, artificial intelligence and computer hardware. AI research is advancing at a tremendous rate. Progress during the last decade has been phenomenal and there is no reason to expect it to slow any time soon. [Page 4 of Project Longshot proposal]

AI is arguably another one of those "not-yet-invented" technologies. AI research may well have been "advancing at a tremendous rate" when this report was created back in 1988 but the sad reality remains that here we are in 2008 , twenty years on from that report, and artificial intelligence remains "not-yet-invented".

To make matters worse, to the best of my knowledge researchers still have only the vaguest of ideas about how to go about creating one. (In fact it could be argued we know more about the principles of creating a functioning fusion reactor or thermonuclear pulse propulsion than we know about the principles for creating a functional artificial intelligence!)

I don't doubt artificial intelligence will be invented some day, but at the same time until the principles are understood nobody will be creating an AI except through sheerest accident.

======
Stephen

[JRehling]

[...]

[Stephen]

As a professional AI researcher, I would hazard to opine that the control portion of a mission like this is not even on the long list of the significant challenges, and is probably way below the level that has already been exhibited by the MERs, for example. If a craft whizzed towards an alien solar system and were able to sense its planets in time to steer towards the interesting ones, I see very little challenge on the AI side.

I'm not an AI researcher, but with all due respect I completely disagree with that statement!

1) For a start who would decide which were the "interesting" planets and which weren't? The researchers back home on Earth before it leaves or the spaceprobe itself on the spot on the fly? If the latter, then that merely begs the question: how often do the MERs themselves decide which is an interesting target and which is not?

If the answer to that is "never" (because the researchers make all those kind of decisions themselves), then that leads us to the next question: could the MERs make such decisions if the researchers allowed them to?

If the answer to that is "no" then how is the sort of decision-making that would be required for an interstellar probe "way below the level that has already been exhibited by the MERs"?

It seems to me you have it the wrong way round! Being able to navigate deftly around a rock (which, as I understand it, is about the level of the MERs capabilities) is several orders of magnitude (at least) below a decision like (for example) deciding that that funny-looking soil with all that white stuff my wonky wheel just dug up is ever so much more interesting than that pretty rock with the nice layering I was making for over yonder.

2) Just how do you define "interesting" anyway? For example, if have a target-rich environment what are the criteria and priorities it would use to decide the few targets it would be able to concentrate its attentions on? Would those be rigidly programmed in prior to launch or would they be a thing the probe would be able to work out for itself based on previous observations it had made? For example, suppose the probe had to choose for a close flyby between a planet that looked like a white billiard ball but with spectroscopic signs of water ice, another which had what looked like craters interspersed by icy plains, a third that was smothered completely in craters, and another that looked like a diseased pizza but with a hint of sulphur dioxide. If that was all it had to go on which would it choose and how would it choose it?

3) Being able to decide which are the interesting targets on arrival is only a fraction of what the computing systems (if you want to call them that rather than "AIs") onboard any interstellar probe will require.

Consider the MERs. They may be able to find their way around rocks, but if in doing so they get caught in a sand trap, even a teensy weensy one like that last time Opportunity got stuck at Victoria Crater, and you rapidly reach the limitations of their intelligence. Instead of being able to work out for themselves how to get themselves out they have to rely on human intelligence back on Earth. Now that is not their fault. The MERs were never designed to get themselves out of those kind of problems. However, those are also the very kind of situations an interstellar probe will need to be able to deal with all by itself. The probe will be completely and utterly on its own. Going into safe mode and waiting for mission control back home to get it out of the fix will not be an option if its two or three light years out.

Granted that some of these problems may be standard ones humans back on Earth may be able to anticipate and work out solutions for it in advance. Others, however, will almost certainly not be. In fact given a mission lasting 100 years, or even 50 years, I would say the latter sort are an inevitability. If the probe gets into trouble it has to be able to diagnose and analyze the problem and devise a solution. If one solution does not work, it has to be smart enough an capable enough to try other alternatives, including ones its makers back on Earth may not have planned for, let alone anticipated. It also has to know how to back out a solution if it doesn't work and which ones can't be backed out of (and so which ones you don't try unless you absolutely have to).

That sort of capability goes way beyond anything the MERs have.

And of course heaven help the probe if should ever get into the kind of life-and-death situation Spirit experienced back at the start of its mission. Yet even that an interstellar probe will need to somehow be able to cope with. If it cannot, then the odds of its surviving to complete its mission will probably be not good, for it will be depending more on sheer luck and good engineering to keep itself out of trouble than having the capability available to get itself out of any trouble it might find itself in.

So assuming that a craft had the ability to determine long-range spectroscopy and light curves of a wide range of worlds in a star system, and had the ability to fly more closely by, say, a quarter of them, I think it would be relatively easy to crunch those numbers. Much harder to get them and to act on them. We would have to forgive the craft if it made suboptimal decisions (are Mars, Titan, and Europa easy to prioritize from tens of AU away?), but it should be able to make decent decisions.

Perhaps.

I can't help wondering though how useful those are going to be. For example, how close would a probe have to be before it could detect (say) nitrogen or oxygen or water in a kind of spectrograms the probe would have available to it--bearing in mind:

1) That the instruments it will be carting along will probably have a limited resolution;

2) The results are not likely to be all neatly available all at once but will accumulate over a period of time as the probe approaches the Centauri system, thus it may not necessarily have all the necessary information until quite late, if at all;

Above I gave an equally optimistic example myself that presumed the probe would have the telescopic capability of the Hubble telescope built in. More realistically, the kind of pictures the probe would likely be presented with (and on which it would have to make its decision) would contain a single tiny point of light a few pixels across (at best) with (at best) a few vague smudges on it. Similarly, the spectrograms may well be composed of some indeterminate squiggles that may well be interpretable in several different ways rather than in one single unmistakable fashion.

In other words, the kind of data it may be presented with to make a decision upon may well be sort that even a trained human being of high intelligence and long experience may well have problems making a competent decision on the basis of. It may also be the sort that may require the kind of lengthy study, special enhancement techniques, and access to a library of reference material that the probe may or may not have available.

BTW, note that thus far I'm making the same assumption other people here seem to be making: that the probe would be a flythrough (albeit Longshot would have been at 5% of c rather than the 10% being assumed above). However, in the particular case of the Longshot probe, the proposal was in fact for an orbiter, not a flythrough; and of Alpha Centauri B at that rather than the more Sun-like Alpha Centauri A. In that context note the study's rationale (p28): "Beta was chosen as the target star because it is a dK-Type star, about which we have very little data, while Alpha is a G2 type star like our own, which we have studied extensively." In other words, the star rather than any accompanying planets would have been the mission's primary target. The probe's instrument package would presumably have been tailored accordingly.

Once in orbiter the probe's manoeuvring capabilities, and thus its ability to do flybys of any promising looking planets, would (presumably) have been limited.

If there's a term that heads one down the wrong path, it's to speak of "an AI", which calls to mind some full-blown human mind that can cry and love and so on. Nobody researching AI talks about "an AI". That was sort of an Omni Magazine, 1980ish idea. Since then, artificial intelligence has aced all kinds of tasks, and some of the things I've seen programs do (even ones I've written) leave me pleasantly surprised.

Glad to hear it. But that is arguably not what most of us non-AI researchers would term "artificial intelligence" (even granted that my own use of "AI" was a mere shorthand because I did not want to keep typing out "artificial intelligence").

That said, it is true my intended interpretation was indeed that of a thinking machine, and thus by implication one with a human-like intelligence. Or so at least I interpreted the Project Longshot proposal's own words: "This fact necessitates that the probe be able to think for itself." (pg 4)

The authors presumably did not use that word "think" lightly.

How many of the systems do you work with or know of could be said to "think for [themselves]"?

I don't doubt that many if not all those systems are extremely capable. They may even have some degree of learning capability. However, I will also hazard a guess that many if not most of those specialise in particular areas; and probably (in general) narrowly defined ones at that. Outside their particular speciality, however, would I be correct in assuming they would (probably) be at sea?

Why such an approach may have its uses on such a probe, especially among its various subsystems, the primary computing system will almost certainly need to have something approaching human intelligence, learning capability, and problem-solving capability. In other words something approximating a human mind, albeit I draw the line at the definition you imply with your reference to "mind[s] that can cry and love and so on". I think you're confusing the human mind with human emotions, many if not most of which arguably have physical causes that little to do with intelligence. You don't need to cry, for example, to exhibit intelligent.

To send out such a probe with a capability less than that would to hobble the mission even before it left the ground, as the Project Longshot team clearly recognised.

Remember, this will not be any MER-style mission with a team of human problem solvers only a few light minutes away. For all practical purposes it will be cut off from human assistance for most of its journey. If it gets into trouble it will have to get out of that trouble all by itself.

Are present-day artificial intelligence systems up to that level?

======
Stephen

[hendric]

Are present-day artificial intelligence systems up to that level?

I think the issue with AI's is a bit of a non-problem. We regularly update software in-flight, why not with this probe? Say you need to start doing your planning 5 years before arrival, and it takes about 10 years to get your SW to the spacecraft, then your spacecraft can run an AI designed up to 85 years in the future to detect,plan, etc for the landing. Just be sure to launch with a seriously over powerful CPU (well, a couple anyways for backups etc obviously) and a ridiculous amount of extra memory.

Plus, the scientists planning the launch will already know the inclination of the system, right? Instead of aiming for the center of the star, the aimpoint could initally be targeting the habitable zone on the side moving away from Earth. Of course, you'd have to aim for where the system will be in the future, since it would move quite a bit in those 100 years.

Given a sufficiently large telescope, the algorithm for choosing which planets to visit should be fairly simple, right? If there is planet with oxygen in the atmosphere, go there. If not, if a planet has liquid water, go there. If not, if a planet has between .2 and 5 Earth masses, go there. If not, if there is a planet within the habitable zone, go there. If not, attempt to flyby as many of the planets as possible. This would miss Titan, but catch Earth, Venus, and Mars.

With knowledge of a Jovian class planet, would it be possible to plan slingshots through to other planets? That might be too dangerous if they have moon systems similar to ours...Or radiation environments...

Anyways, the AI issue should be easily tractable. The bigger issue is how much mass you expect to send. At a minimum you'd need a nuclear reactor powered spacecraft of some sort, a large telescope for planning the visit, and a giant radio antenna for communicating back home (or maybe a laser comms). And redundant on top of redundant systems. Maybe 100+ kilo kg? (10x Hubble)

This all raises an interesting question: If someone sent a similar probe through our system, would we be able to detect it? Or would it zip by completely unnoticed? If the homeworld is in constant communications, we might see the beam once the spacecraft and sun are sufficiently aligned. But most likely, once the software is updated, the homeworld will be listening for any results instead of transmitting...I think we probably wouldn't notice an interstellar probe, unless it is of the giant-RAMA-spaceship type...

[Stephen]

I think the issue with AI's is a bit of a non-problem. We regularly update software in-flight, why not with this probe? Say you need to start doing your planning 5 years before arrival, and it takes about 10 years to get your SW to the spacecraft, then your spacecraft can run an AI designed up to 85 years in the future to detect,plan, etc for the landing. Just be sure to launch with a seriously over powerful CPU (well, a couple anyways for backups etc obviously) and a ridiculous amount of extra memory.

1) What constitutes a "seriously over powerful CPU" and a "ridiculous amount of extra memory"?

Think back to the computing hardware that is currently flying in the Voyagers and you will see how that is not a serious proposition! What would have seemed like a "seriously over powerful CPU" and a "ridiculous amount of extra memory" back in the 1970s is now a mere drop in the proverbial bucket required by much of the software now running on the computing hardware of 2008, let alone the kind likely to be available in the 2060s (1970s + 85 years...), even granted that much of that power in current software tends to be used to display increasingly sophisticated graphical GUIs.

2) I doubt a single CPU, no matter how "seriously over powerful", will be able to run an AI. The human brain seems to make massive use of parallel processing to do the same job, so I suspect (just MHO) the hardware required to run an intelligence of an artificial kind will probably need to do the same as well. The upside: the probe's designers will not necessarily need to include "seriously over powerful" CPUs. Less sorts would probably suffice. The downside: they may well, however, require a very large number of such CPUs, which in turn may tax the power requirements of the probe (amongst other things).

3) As for updating software in flight, sure it's done now although I doubt if anyone's done what you seem to be contemplating: adding the AI **after** launch rather than as before (as part of the original design).

I am however wondering though what the data rate would be for the probe at a distance of 2 or 3 light years from Earth and the size of the update. I notice the Project Longshot paper says: "A laser with an input power of 250 kilowatts would allow for a data rate of 1000 bits per second at maximum range." If we take that as our starting point, then assuming 8 bits per byte (and ignoring the issue of other bits like parity bits etc consumed in the transfer), that equates to about 125 bytes/second, 450,000 bytes per hour, or 10.8 megabytes per day. A software update of about 1 gigabyte would therefore take about 100 days to upload.

Of course this would pale beside the timespan required if the probe, having received the update, then had to verify the uploaded software with Earth (to ensure it was received correctly). If the probe by then was (say) 3 light years out it would be six years (+ another 100 days) after Earth uploaded the last byte that the first of the data to be verified came back. Then a further 3 years before the "OK to proceed" signal could be sent.

In other words, a single 1 gb update could take over a decade to accomplish. (3 years + 100 days to upload update, 3 years +100 days to verify, 3 years to send "OK to proceed".)

Even if the verify were not needed (because the uploaded data was self-verifying) the time involved would not be that much shorter unless the probe did not require Earth's approval to implement the update.

(Of course all of the above assumes that everything uploaded correctly and did not have to be re-sent!)

Given a sufficiently large telescope, the algorithm for choosing which planets to visit should be fairly simple, right?

Are artificial intelligence algorithms "fairly simple"?

With knowledge of a Jovian class planet, would it be possible to plan slingshots through to other planets? That might be too dangerous if they have moon systems similar to ours...Or radiation environments...

Not to mention time consuming. I seem to remember it taking an awful long time for Galileo, Cassini, et al to get around using only gravitation slingshots.

Besides, who will be planning such slingshots? The experts back on Earth or the AI on the probe?

======
Stephen

[djellison]

3) As for updating software in flight, sure it's done now although I doubt if anyone's done what you seem to be contemplating: adding the AI **after** launch rather than as before (as part of the original design).

Whilst personally, I consider this entire thread to be more about science fiction than anything else (check the forum rules for what that means) - I will pick you up here.

The MER's have had several significant flight software updates that have given them abilities not even considered before launch. It depends where you definition of AI lies - but go-and-touch, automated DD/Cloud watching, D-star etc, are all abilities that were developed long after landing and uplinked to a vehicle designed, built and launched without those abilities in mind. It's not unreasonable to say that the MER's that drove in their primary mission lacked intelligence (but not autonomy), and the MER's today have both an element of intelligence and autonomy.