Skip to content

Clause learning against propagation: a negative result

Clause learning gave no growing lead over plain arc-consistency propagation on wrong prefixes of the real pieces: the pre-stated stop line fired on summed work, and in counts learning is worth 0 bits over arc consistency. The toy boards (5x5 to 7x7) lean the same way but do not establish it. Nothing here is Eternity II progress.

by Raphaël Anjou

experimentdead endmeasuredrefutedExact methodsReduce the searchBacktrackingUpdated 2026-11-14
Reproducedeterministic: reproduces byte-for-bytere-verifies a stored boardBudget: Real-piece part: 800 runs, 9.15 core-hours, 1,200 s limit per run. Toy part: 3,500 s limit per run, one thread.
Hardware & run
Native runCPU only
0.56core-hours
Cores
10
RAM
32 GiB
GPUs
0
Machine
10-core Apple M4 laptop, 32 GB; 8-core Apple M1 laptop, 16 GB; 14-core low-power x86 laptop
Budget
Real-piece part: 800 runs, 9.15 core-hours, 1,200 s limit per run. Toy part: 3,500 s limit per run, one thread.
Start
toy boards seeds 1 to 10 per size; real-piece prefixes seeded 1000*depth + index, the same for every solver
Verify the stored board

deterministic: reproduces byte-for-byte

This verifies the board, not the search that produced it.

just experiments learning-vs-propagation

Reruns the analyses and their independent recomputations on the committed raw outputs (standard-library Python, a few seconds). It does not rerun the solver jobs: 9.15 core-hours for the real-piece part, hours more for the toy part.

Keep exploring

Referenced by

Page sourceView as Markdown