Evolution of AI Interplanetary Trajectories Reaches Human-Competitive Levels

New submitter LFSim writes "It's not the Turing test just yet, but in one more domain, AI is becoming increasingly competitive with humans. This time around, it's in interplanetary trajectory optimization. From the European Space Agency comes the news that researchers from its Advanced Concepts Team have recently won the Gold 'Humies' award for their use of Evolutionary Algorithms to design a spacecraft's trajectory for exploring the Galilean moons of Jupiter (Io, Europa, Ganymede and Callisto). The problem addressed in the awarded article (PDF) was put forward by NASA/JPL in the latest edition of the Global Trajectory Optimization Competition. The team from ESA was able to automatically evolve a solution that outperforms all the entries submitted to the competition by human experts from across the world. Interestingly, as noted in the presentation to the award's jury (PDF), the team conducted their work on top of open-source tools (PaGMO / PyGMO and PyKEP)."

Comments

In other news...

[14erCleaner]

John Henry won the battle, but lost the war. How is being outcalculated by a computer news? Just because it's a hard problem?

Re:In other news...

[Rich0]

John Henry won the battle, but lost the war. How is being outcalculated by a computer news? Just because it's a hard problem?

It involves a certain amount of intuition. You'd always use a computer to optimize a trajectory, but picking the overall approach to be used requires some educated guessing. You can visit any other planets along the way, or their moons, and you can visit them at an optimal point along their orbit or somewhere that is non-optimal (from the standpoint of that particular encounter). You can launch today at one cost, or wait 20 years and maybe launch at a cheaper cost.

So, the current approach is generally to have physicists come up with a couple of basic plans, then use computers to optimize each one, and then see which works best, or perhaps iterate.

Looking at it another way, this is similar to any other problem where you're trying to find the lowest minima in a function that has many local ones. Finding the nearest minima is easy - finding the best is much harder.

Re:In other news...

[Rich0]

Yeah, but there's only so many bodies out there in the solar system. Probably under 200 planets/moons/rocks out there that can be used for this purpose. It probably wouldn't take much to throw a super-computer at the problem and just crunch through all the possibilities.

Naively I'd be inclined to agree, but there are a lot of combinations. Cassini encountered Venus twice, the Earth once, and Jupiter once on the way to Saturn. Jupiter is a pretty natural target to hit once at the end of the trip since it is so large and far out, and you really don't want to hit it more than once unless you want to have your grandkids crunching your data when the mission is over. However, for the inner planets there are a lot of possibilities. Again, you don't have to hit them at your periapsis even if that is where you get the biggest boost - being able to squeeze in one more non-optimal boost is better than none at all, or you might just be using the encounter to change your orbit to make up for a previous encounter sending you off in the wrong direction.

If you're not just looking for encounters at periapsis then there are many combinations and a continuum of variations on each. The fact that the guys at NASA are still doing much of the high-level planning by hand speaks to the difficulty of the problem.

Disclaimer - I'm not an astrophysicist, but I have on occasion had fun playing with Orbiter.

Re:Isn't this already done by computers?

[K. S. Kyosuke]

This is about the boundary conditions. It's like the difference between using TeX to typeset a document letter by letter (= repetitive trajectory integration steps) versus using the same computer to compose a meaningful document to be printed in the first place (= designing the trajectory and mission profile creatively).

Re:Isn't this already done by computers?

[h4rr4r]

Obviously you manually provide some limits.
Humans do not start from nothing each time they calculate these either.

Not human vs AI, human vs human competition

[slew]

The competition was not that AI was in competition with humans to develop spacecraft trajectories, it was that humans were in competition with other humans to quickly develop frameworks create the best mission design in a complicated search space that had multiple local optima and unusual constraint functions (preventing the use of "canned" solvers).

One of the critera used to select the problem was...

Problem is easy enough to tackle in a 3-4 week timeframe for experienced mission designers or mathematicians, including exploration of new algorithms.

Of course many of the teams in the competition probably used AI-like frameworks to find the actual trajectories so it's unsurprising an AI technique won. Although perhaps some teams tried other non-AI-like searching techniques (like pseudo-objective functions), I'm pretty sure none of the teams chose to use human pondering to come up with mission designs.

Re:Isn't this already done by computers?

[K. S. Kyosuke]

This is not about limits. You can only bruteforce stuff that is finite in size. Anything in R^n has an infinite number of alternatives and therefore trying all of them (which is what "brute force" means) is patently impossible.

Re:Child's play compared to the tough questions

[newcastlejon]

Call me when they can answer the really tough questions, such as, "Does this dress make me look fat?"

Interesting game. The only winning move is not to play.

Re:Isn't this already done by computers?

[NeutronCowboy]

Drop me a line when a computer is better than humans at something human brains are actually good at.

Like my old AI prof said: When we don't know how to do it, it's AI, when we know how to do it, it's Engineering. Similarly, when we don't know how to write a program that is better than a human at something, it's because human brains are very good at it. Then, when we do know how to do it, it's something that human brains really aren't very good at it. You can easily see this by how much smaller the space of "things the human brain is good at" has become. It wasn't so long ago (say, 1985), that Chess was seen as an excellent litmus test of AI. Now it's seen as little more than a beginning software engineering project. Heck, it wasn't so long ago that people were pegging Jeopardy as an excellent litmus test of AI, what with the puns, the historical cross references and the pop culture in it. Now a computer has soundly beaten the best Jeopardy players ever.

I predict that we will identify strong AI only when it has already enslaved us (queue the joke that we're already slaves to computers).

Re:Isn't this already done by computers?

[gl4ss]

they have some limits on where they want to end up and where they can launch from(when is a variable though that needs to be taken into account).

you can brute force it just fine, not in the sense of checking every possibility of course, but by checking enough possibilities. that's what I suspect the evolution is here anyways, evolving the possible paths likely to be good, so if some branch seems like no tweaking of it could ever provide a good answer then abandon that branch of the family tree. so even though their algos couldn't provide you with something they can say for certain is 100% optimal answer it's still competitive with the human chosen trajectories.

So.. the question would bickering about if evolution is brute forcing(it is, but the parameters tend to change).