Thread by @Thomassaf

With all the rage about ChatGPT, have you ever wondered about the differences between AI and more classical technologies? Join me for a brief tour, with its relation to Aerospace. @RedDivulga, @filarramendi, @uc3m y @doctoradoUC3M.  @PhDAaeroUC3M #PhD #Aerospace #Engineering

ChatGPT, and its equivalents work like buttonsmashing. You have a bajillion buttons, and them you mash, tweak, and tune them randomly. Like this young Alien Stu is doing in the Pixar short movie Lifted.

Stu is trying to learn how to abduct humans. I know, a very Aerospace oriented scenario. Consider him as Chatgpt in its beginnings. He pushes random buttons, and not surprisingly, nothing really works. However, the more he tries, the more he learns the buttons

After many attempts, eventually, he will learn how to use all the bajillion buttons just like his mentor Mr.B. It is in many ways how most of us learns; You try, you get some feedback (Failure or success), and then you try again with your newly acquired experiences.

Now, that’s AI / ChatGPT. And some of the problems with it is that you need a computer powerful enough that can tune/tweak millions upon millions of buttons. In doing so, it needs to try trillions upon trillions of times to gain enough experience.

Just recently computers have become powerful enough to be able to this, and that is why AI has suddenly become so popular and functional. However, being able to train in real scenarios trillions of times is not exactly something one can do in many domains, like Aerospace.

Why don’t we just simulate them? Well, space is difficult, especially places we haven’t been before. It is challenging to accurately create a simulation that perfectly catches everything an asteroid can throw at you, figuratively and literally. So, what do we do?

Something I have been working on is an approach that is called Model Predictive Control, or MPC for short. In some ways, it behaves very similar to AI as we have previously discussed, in others, it is quite different. So, let us break down in detail (twitterstyle) how MPC work

Model implies that we have a model of the environment. Predictive means that we use that model to predict what is going to happen in the future. Finally, control means that we apply our thrusters such that we end up where we have predicted we are going.

“But” I hear you say, “Didn’t you just say that simulations and models are difficult to make?”. Well, yes. And it doesn’t help that the models we can use in MPC have to be super-duper simple. But, as opposed to AI, we “know” that our models are bad, and can do something about it.

Instead of Bajillion buttons to tweak, we take a simple model with 10s of buttons. We try it in our simulations a couple of times, maybe even compute some fancy pancy solutions, and estimate how bad it is. If we know how bad it is, then it is not that hard to compensate for it.

As an example, imagine you are driving a car/spaceship with a loose steering wheel. After some quick learning, you will be able to atleast keep the car on the road /star-lane. In essence, that is the principle of some of the most advanced space driving techniques. Like this guy:

In general, AI is like a chef with a million spices, trying to create the perfect dish. MPC is like a minimalist chef who only uses salt and pepper, but knows exactly how much to use to make it taste great.

If you want the full image of the equations (you dont), they are here. They more or less do exactly what we have laid out before. Model, Predict, and then compute the control. We do this as often as we can, and BOOM, you are landing on asteroids, without those pesky AI’s.

Well, maybe not BOOM. Actually, hopefully not BOOM, unless you are the United States of A’, which performs its speciality in purposefully going Boom on asteroids. Check out the DART mission of NASA for a taste of their love of explosions.

So, what is best? Well, AI can be best for a doing stuff we have done before. As it can tweak its many buttons perfectly to match all the training data. However, atleast for the current time, systems with few buttons carefully tuned is one of the safest ways to explore the beyond

#AI #MachineLearning #DataScience #Robotics #Aerospace #SpaceExploration #STEMeducation #ScienceCommunication #Aerospace #education #PhD #phdlife