I am learning reinforcement deep learning and working on applying it to the rather difficult problem of routing PCBs. It's a classic NP hard problem where you're trying to a find a set of 3d paths (x and y are continuous, z is discreet) that fulfill arbitrary constraints like length, distance to other paths, etc.

I believe that various forms of deep learning can enable leaps in autorouting because computers are far better suited to combinatorial optimization than humans and learning algorithms are finally starting to emulate a level of intuition that makes routing lots of traces (terminology for a connection on a PCB) relatively easy for the human brain. I also think that reinforcement deep learning is the ML algorithm best suited to this problem because it mirrors how humans route: in discreet steps that can be encoded as actions like go up 10mm, go down one layer, etc, guided by visual feedback and a design rule checker.

I'd greatly appreciate any advice or papers/projects that show how to use reinforcement learning for problems different from playing a video game. What I need is a significant deviation from OpenAI gym solutions and without a lot of experience in machine learning, it's hard to know where to look and what terminology is relevant. I'm jumping into the deep end here so the best advice is to start with learning linear algebra and simpler neural network architectures but starting with a hard, relevant, and exciting problem is how I learn a new subject best with limited time and conceptually deep learning just "clicks."