Comment on i hate myself and i want to die lol
fossilesque@mander.xyz 2 days agoMy environmental proxies have been a conundrum for machine learning (I’m pretty sure why, writing a grant on that), so I can sympathise. I have some tools for you, if you’d like for me to share. Dm me and I’ll send you a bunch of stuff when I’m less inebriated.
PM_ME_VINTAGE_30S@lemmy.sdf.org 2 days ago
Re: thread about your PhD manuscript. Strongly recommend waiting until not shit faced 😆
Short answer: send whatever you have
Long answer:
So I’m taking a course on advanced machine learning. I know pretty well how to use SkLearn, Pytorch, and HuggingFace transformers to implement machine learning algorithms.
But the thing is that my research is in dynamical systems, specifically power system dynamics. (I’m the “control guy” in my group, i.e. I know more about math (specifically control theory) than I do power systems.)
So my group is really interested in getting theoretical guarantees because … well, ultimately we’re literally trying to keep the lights on. So I’m taking this machine learning course to learn about asymptotic analysis and finite-sample analysis of the convergence of ML algorithms, i.e. to mathematically analyze machine learning algorithms. Power systems are extremely high-dimensional and nonlinear, and as more wind and solar PV plants are added to the grid, we’re actually going to have to change how we control the grid because they don’t “act like” synchronous machines (i.e. hydro, fossil, nuclear, biodiesel plants).
And to the ML professor’s credit, he has taught these things and he’s clearly very careful about the mathematics. But he really does over-rely on LLMs which … I’m having trust issues.
So one thing our group does is data-driven analysis of dynamical systems. For controlling or observing a dynamical system (like a power grid) our group has been looking into the Koopman operator framework. The Koopman operator is a composition operator that converts a nonlinear dynamical system into an infinite-dimensional linear dynamical system. Unlike standard linearization, this method makes no approximation. Now, by approximating this infinite-dimensional linear operator with a finite-dimensional linear operator, i.e. a matrix. Then in this framework, we can do data-driven control (take the system to an arbitrary state), estimation (get the internal state, for example see if the system is in a dangerous state), and identification (for a mechanical example: you know that you have a pendulum and you have recorded the trajectory of the pendulum, but you want to know the mass and length of the pendulum).
So we have a huge interest in machine learning and it’s impressive results, but we are also going be honest with whatever theoretical guarantees we can prove. So yeah if you have any literature on machine learning, specifically supervised and reinforcement learning, please send.