Initial Problem:
Take Dijkstra’s Algorithm. It’s a great algorithm, but due to the nature of the problem, it’s really inefficient as the problem (graph size) grows larger. In fact, the time complexity is , where is the number of nodes you have.
Mind Blowing Part:
What if you trained a neural network on the inputs and outputs of a bunch of examples (so start point, end point and the path taken). Then (assuming the model learned something), you have a model that approximates the best path, but at …
Credit to Oliver Groth, who taught me this idea over a pint lol