Hacker News new | ask | show | jobs
by wanderingjew 2398 days ago
> This is an unsolved problem...

Yes, this is the point. It's NP-Hard. If you solve this problem, you can make far, far more money doing something besides routing mechanical keyboards and Internet of Things sensor cruft.

Autorouters have been in development for the last fifty years, starting with wire-wrap machines at Digital, and going on to the work of very, very smart people at Altium and Autodesk. The smartest people in their field have been working on autorouters for decades, and this company wants to solve it with 'the cloud' and 'AI'. Sure, buddy.

150 pairs and 2 layers is abysmally limited for anything but the lowliest hobbyist (read: poorly designed) boards, and there are no examples whatsoever of what this product produces. Like, really, great job for producing a demo to show to investors but you might also want to demonstrate your demo.

Oh, and if you're using machine learning on PCB design, that means you need to train your models somehow. That means your training data is absolute crap, because most designs for Open Source hardware are objectively crap. You would be better off paying someone in China $40 to lay out your board, which would also have a 24-hour turnaround. Which brings me to my next point...

2 comments

I just don't understand how deep learning is amenable to this problem. It's just a straight up optimization problem. How is deep learning appropriate?
I think it is probably appropriate. Chess and go are just "straight up optimization problems", but they're too difficult for traditional optimisation algorithms to work. You need something to do some fast pattern recognition to cut down on the search space. This is similar.

I expect if you search the literature you'll find a ton of work on this.

Unless you believe humans are solving NP-hard problems somehow, it seems reasonable that we learn heuristics to solve specific problems. Theoretically, so could an ML model, and indeed, if you look at recent results, using things like graph NNs on traveling salesmen problems have done very well.