Sample Efficient Evolutionary Algorithm for Analog Circuit Design
bair.berkeley.edu
bair.berkeley.edu
We've been working on an internal project for layout of digital circuits using stochastic search and ML and have been having good results. ML and AI in general will likely have a pretty big impact on circuit design in the near future, excited to see other work in this area ^_^
Because:
1. They're working on the cutting edge of miniaturization and autoroute functionality just doesn't cut it at that level. You need a human understanding of the end product as in consumer electronics you have to plan in advance several (often for lower cost) revisions of the same product for the future.
2. All the ML/AI experts are expensive and they have much lucrative fish to fry such as getting you to buy or click on things rather than optimize PCB aoutoroute for some eCAD company that can't justify their expense as most of their customers will still route by hand since the hardware business is very resilient to change.
3. PCB layout is done(in Western Europe at least) mostly by technicians, not engineers, and their labor costs are cheap as technical high-schools churn hundreds every year. You don't need a university degree and sometimes the company will pay your training for the eCAD tool they're using. In my area, a city with lots of hardware industry, it's basically a blue collar job.
Source: FW dev in the consumer electronics business
Perhaps some kind of GAN approach could be used to force an optimizer to perform autorouting with a bias towards "human-like" circuits that actually readable and understandable.
JITX is tackling the PCB problem however.
"In our proposed method, we devised a model to predict the performance before simulation and only simulate those samples which have better predicted performance."
In other words, build a model for the system, use that for (cheap) sample point evaluation, occasionally check (doing an expensive circuit simulation) and adjust.
P.S. back in 80s the guys doing simulated annealing were not claiming to be AI researchers, and the actual AI researchers were doing other stuff (neural networks and knowledge-based systems, among others).
(Not so much NN in the 80s except for a few holdouts like Rumelhart and Hinton).
You could imagine a GA as being a multiple simulated annealing algorithm in which you occasionally cross pollinated your solutions.
Technically, in GA you keep the top solutions and SA you will move to a less optimal solution with some probability. But the mechanics of them are quite similar.
[0] https://static.aminer.org/pdf/PDF/000/308/779/an_evolved_cir...
This makes me wonder, what if the algorithm finds a patented circuit, would it be allowed?
Another thing I wonder is if instead of starting with a fixed topology, you start with a complete graph, would the algorithm automatically find the correct topology (where e.g. resistors that are dropped simply become very large). Or would it get stuck in a local minimum?
Back in time around year 2000, Scientific American had published an article about using genetic evolution algorithm to create circuits. 20 years later, the same idea keeps pop up again and again, but no commercial product has really claimed they were generate by algorithm.