Anybody got some ideas to optimize today? I’ve got it down to 65ms (total) on my desktop, using A* with a visitation map. Each cell in the visitation map contains (in part 2) 16 entries; 4 per direction of movement, 1 for each level of straightaway. In part 2, I use a map with 11 entries per direction.
Optimizations I’ve implemented:
- use a 2D array instead of a hashset/map. No idea how much this saves, I did it in the first place.
- the minimum distance for a specific cell’s direction + combo applies for higher combo levels as well for part 1. For part 2, if the current combo is greater than 4, we do the same*. Gains about 70(!!) ms
- A* heuristic weighting optimization, a weight of about 1% with a manhattan distance heuristic seems to gain about 15 ms (might be my input only tho)
*Correctness-wise: the reason we’re splitting by direction is because there’s a difference between being at a cell going up with a 3 combo but a really short path, and going right with a 0 combo but a long path. However, this is fine because a 3 combo in the same direction as a 0 combo is identical, just more restrictive.
Optimizations that could be done but I need to ensure correctness:
the same optimization for the combo, but for directions. If I’m on a specific combo+direction, does that imply something about the distance for another direction? Simply doing the same for every non-opposite direction isn’t correct
Code: https://codeberg.org/Sekoia/adventofcode/src/branch/main/src/y2023/day17.rs
Warning: quite ugly, there’s like 8 copy-pastes for adding to the queue
I did a different heuristic for A*: I ran a Dijkstra backwards from the bottom right corner and computed the heat losses for each block if there were no movement restrictions whatsoever. Maybe that’ll help shave more milliseconds :)
