coorl-lost-cities/experiments/traversal_policy_boundary/README.md

Traversal Policy Boundary Microbench

Purpose: separate the current traversal hot path into game/encoding overhead, single-request policy boundary overhead, and batched forward lower bounds.

Run:

uv run python experiments/traversal_policy_boundary/bench_policy_boundary.py \
  --traversals 32 \
  --runs 5 \
  --warmup 1 \
  --corpus-size 512 \
  --component-repeats 4 \
  --forward-repeats 32 \
  --device cpu

CUDA spot check:

uv run python experiments/traversal_policy_boundary/bench_policy_boundary.py \
  --traversals 8 \
  --runs 3 \
  --warmup 1 \
  --corpus-size 512 \
  --component-repeats 2 \
  --forward-repeats 32 \
  --device cuda \
  --output experiments/traversal_policy_boundary/results_cuda.json

2026-05-07 results, legacy/deep-cfr/configs/default.yaml, RTX 3090 host:

Device	Component	Median us/call	p95 us/call
CPU	encode + legal	3.10	3.81
CPU	push + pop	0.15	0.22
CPU	policy boundary bs=1	111.50	125.46
CPU	torch forward bs=64	12.84	13.14
CUDA	encode + legal	3.16	3.88
CUDA	push + pop	0.16	0.25
CUDA	policy boundary bs=1	181.30	194.77
CUDA	torch forward bs=64	2.55	2.75

Interpretation: the game mechanics and state encoding are not the dominant cost. The current one-row policy boundary dominates, and CUDA only becomes attractive once requests are actually batched.