Hello all,
For the last couple weeks I've been working on
robodoan, an automated solver for 3^4 using incremental blockbuilding. Although it can't yet do a full solve, it's able to reliably compute a 60-75 STM solution to the first two layers in under 10 seconds on my laptop. For comparison, the previous shortest solution by Charles Doan (188 STM) completed F2L in 114 STM. It works by pairing blocks one at a time, brute-force searching ahead by up to 5 moves (usually ≤3) to pair blocks using some heuristics. To get short solutions, it tracks up to ~10,000 possible solutions at once and picks the best ones at each step. See
the README for more details.
Attached is a 125 STM solve Luna Harran and I did in 4 stages:
- F2L (63 STM, automatic) - We generated 5 scrambles using MC4D, loaded them into robodoan, and manually reviewed several of the generated solutions. The shortest F2L solution was 61 STM, but it had a poor OLC. We found one that was 63 STM and had a decent OLC (orient last cell).
- OLC (21 STM, manual) - We solved OLC using 2 algorithm executions, including a slice RKT-canceled fat sune and a 6-move 4D OLC algorithm.
- 2cPLC (8 STM, manual) - We solved 2cPLC (permute last cell) using an adjacent 3-cycle that canceled into the last OLC algorithm.
- PLC (33 STM, computer-assisted) - We entered the final cube state into Cube Explorer to get a 3D solution and then manually found a decent RKT-cancel for it.
The MC4D log file shows 150 moves because MC4D does not allow 180-degree ridge turns and because there are several corner & edge twists that I didn't bother simplifying. To verify the move count, load the .hsc file in
Hyperspeedcube.
I've lost motivation to continue working on the program for now, but in the future I'd like to come back to it and implement searching for OLC and PLC. I expect it'll then be capable of finding solutions less than 110 moves, possibly as low as 100.