Records
Expectimax samples-by-depth sweep against greedy
Registration 1126bd34-1166-4de9-934d-e6274a23b622 (study expectimax/samples-by-depth), run 2026-09-18 on AMD Ryzen AI Max (Strix Halo), 32 threads, 125 GB unified memory, Radeon 8060S via ROCm, Linux 6.18. Server commit 46ec303, research commit 097c923 (registration files as committed; this record added afterwards).
Command, once per cell with samples in 4, 16 and 64 and depth in 1, 2 and 3:
cd server && cargo run --release -p armies-arena --bin matchup -- --red 'expectimax:{"samples":S,"depth":D}' --blue greedy --games 100 --seed 0 --rules training --threads 32 --moves --out ../research/runs/1126bd34-1166-4de9-934d-e6274a23b622/sS-dD.json --record-dir ../research/runs/1126bd34-1166-4de9-934d-e6274a23b622/records/sS-dD --tag 1126bd34-1166-4de9-934d-e6274a23b622,sweep
cd research && python3 scripts/paired_contrast.py runs/1126bd34-1166-4de9-934d-e6274a23b622/sS-dD.json| Worlds | Depth | Wins | Draws | Losses | Effective | 95% Wilson | Contrast (95%) | Mean plies | ms/move | Nodes/move | Wall s |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 4 | 1 | 57 | 1 | 42 | 0.575 | 0.477 to 0.667 | 0.15 (-0.07 to 0.37) | 128 | 0.20 | 420 | 0.1 |
| 4 | 2 | 53 | 0 | 47 | 0.530 | 0.433 to 0.625 | 0.06 (-0.17 to 0.29) | 138 | 2.25 | 5,207 | 0.7 |
| 4 | 3 | 58 | 2 | 40 | 0.590 | 0.492 to 0.681 | 0.18 (-0.03 to 0.39) | 137 | 18.33 | 44,103 | 5.5 |
| 16 | 1 | 59 | 4 | 37 | 0.610 | 0.512 to 0.700 | 0.22 (0.00 to 0.44) | 136 | 0.80 | 1,630 | 0.2 |
| 16 | 2 | 66 | 3 | 31 | 0.675 | 0.578 to 0.759 | 0.35 (0.13 to 0.57) | 151 | 9.50 | 21,594 | 3.1 |
| 16 | 3 | 65 | 0 | 35 | 0.650 | 0.552 to 0.736 | 0.30 (0.07 to 0.53) | 140 | 71.29 | 157,816 | 19.5 |
| 64 | 1 | 45 | 12 | 43 | 0.510 | 0.413 to 0.606 | 0.02 (-0.20 to 0.24) | 144 | 2.66 | 5,797 | 0.8 |
| 64 | 2 | 62 | 1 | 37 | 0.625 | 0.527 to 0.714 | 0.25 (0.02 to 0.48) | 143 | 32.29 | 77,727 | 10.5 |
| 64 | 3 | 61 | 2 | 37 | 0.620 | 0.522 to 0.709 | 0.24 (0.01 to 0.47) | 142 | 147.08 | 333,394 | 50.5 |
Endings over all 900 games: flag captured 850, no moves 25, quiet limit 25.
Decision rule
Depth claim (at 16 worlds, the depth-3 interval above the depth-1 interval): depth 1 is 0.512 to 0.700, depth 3 is 0.552 to 0.736; not supported. Worlds claim (at depth 2, the 64-world interval above the 4-world interval): 4 worlds is 0.433 to 0.625, 64 worlds is 0.527 to 0.714; not supported. Browser candidate (strongest cell under 200 ms per move, ties to the cheaper cell): 16 worlds at depth 2, 0.675 effective at 9.50 ms per move.
Provenance
Archived JSON under runs/1126bd34-1166-4de9-934d-e6274a23b622/ (sS-dD.json, 31.9 MB in total, every game with its move list and per-ply detail) and the per-cell record folders under runs/1126bd34-1166-4de9-934d-e6274a23b622/records/. SHA-256 of each cell file: s4-d1 4c00d380743ee19d; s4-d2 2a3b12b8c683ae8e; s4-d3 f8cc2a6ed13899ee; s16-d1 4a70e4d76e8211f0; s16-d2 feabcb1eeebe97a9; s16-d3 4b007751cfc64786; s64-d1 93ce67e8ff4dcd99; s64-d2 cb30fee9aed50160; s64-d3 57aa04dc6a729d38.
Games planned 900, attempted 900, completed 900, interrupted 0. The curated replay library holds seed 0 from both seats for every cell (18 records under records/games/expectimax-sweep/), as the registration states. No deviation from the registration; every cell stayed under the 2,000 ms per move budget.