Records

Belief rollouts against expectimax over sixteen worlds

Registration 6e83e965-97e5-4171-a761-dbbdcea3fa01 (study expectimax/rollout-rung), 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:

cd server && cargo run --release -p armies-arena --bin matchup -- --red rollout --blue 'expectimax:{"samples":16,"depth":2}' --games 100 --seed 0 --rules training --threads 32 --moves --out ../research/runs/6e83e965-97e5-4171-a761-dbbdcea3fa01/rollout-vs-expectimax.json --record-dir ../research/runs/6e83e965-97e5-4171-a761-dbbdcea3fa01/records --tag 6e83e965-97e5-4171-a761-dbbdcea3fa01
cd research && python3 scripts/paired_contrast.py runs/6e83e965-97e5-4171-a761-dbbdcea3fa01/rollout-vs-expectimax.json
PlayerWinsDrawsLossesEffective95% WilsonMean pliesms/move
rollout190810.1900.125 to 0.2781445.860
expectimax810190.8100.722 to 0.8751448.216
ending: flag_captured100

Per-seed paired contrast for rollout (score as Red plus score as Blue minus one): mean -0.620, 95% normal interval -0.813 to -0.427 over 50 seeds (6 positive, 7 split, 37 negative). Rollout won 10 as Red and 9 as Blue. Mean playouts per rollout decision 681. Wall time 4.1 s on 32 threads.

Decision rule

The hypothesis (rollout does not beat expectimax at equal or lower cost) holds: the contrast interval is -0.813 to -0.427 and rollout spent 5.86 ms per move against expectimax's 8.22 ms. The rollout rung is not promoted.

Provenance

Archived JSON: runs/6e83e965-97e5-4171-a761-dbbdcea3fa01/rollout-vs-expectimax.json (SHA-256 64577fc3d9f2c337fe7994dec26bb73bea24b2fc72c6693c3bd104b30a3476a4), 3.9 MB with every game's move list and per-ply detail, so it is not copied next to this record. Games planned 100, attempted 100, completed 100, interrupted 0. The curated replay library holds seeds 0 to 4 from both seats (10 records under records/games/rollout-vs-expectimax/), as the registration states. No deviation from the registration.