知识库收录材料 · 原始文件 ↗ · 来源、版本与许可

APPO Basic versus Jev

All three public APPO checkpoints completed 48/48 ViZDoom Basic cases when their original training rendering was restored. Each scored 6/6 on test and 6/6 on OOD, compared with Jev's 3/6 on each split. These checkpoints are practical candidates for generating Basic action demonstrations.

The rendering configuration was decisive in this experiment. The initial benchmark used NanoJev's existing framebuffer settings. We then checked the checkpoint's original source, restored its training rendering, verified all three policies on development cases, and ran the complete fixed case set. Both rounds are retained below.

Results with the training rendering

All controllers in this table use greedy selection with 10% uniform exploration and sampling seed 17. Test repeats each action for up to four ticks; OOD repeats for up to eight. Native termination remains enabled. A successful episode has a positive kill-count increment.

Controller	Train	Dev	Calibration	Test	OOD
Jev	10/24	3/6	4/6	3/6	3/6
APPO 1111	24/24	6/6	6/6	6/6	6/6
APPO 2222	24/24	6/6	6/6	6/6	6/6
APPO 3333	24/24	6/6	6/6	6/6	6/6

Controller	Test mean ticks	OOD mean ticks	Test mean ammo consumed	OOD mean ammo consumed
Jev	167.33	235.67	12.00	16.67
APPO 1111	21.50	25.00	1.17	1.00
APPO 2222	17.67	25.00	1.00	1.00
APPO 3333	17.67	37.00	1.00	1.50

These are means over every episode, including Jev's failures. They measure game time, not inference latency. Each split has six environment cases, reused across the three checkpoints; the three models do not create eighteen independent test scenarios.

The map, case seeds, buttons, action durations, and success criterion are shared. APPO consumes RGB frames with recurrent memory; Jev consumes structured visible-object boxes, player variables, and observation history. This is a comparison of complete controllers with their respective input representations. Predict Position is not covered by these Basic checkpoints.

Machine-readable results: training-rendering follow-up.

Initial benchmark with the existing rendering

The protocol fixed checkpoint 1111 with greedy selection and epsilon 0.1 as the primary comparison before APPO outcomes were inspected. Its initial test result tied Jev, and its OOD result was lower. All nine planned runs completed.

APPO checkpoint and controller	Test	OOD
1111 greedy, epsilon 0.1 — primary	3/6	1/6
1111 greedy, epsilon 0	3/6	0/6
1111 sampled	2/6	1/6
2222 greedy, epsilon 0.1	4/6	3/6
2222 greedy, epsilon 0	4/6	3/6
2222 sampled	4/6	3/6
3333 greedy, epsilon 0.1	5/6	3/6
3333 greedy, epsilon 0	5/6	3/6
3333 sampled	5/6	5/6

Complete initial results include every split, metric, source hash, and paired case comparison. The training-rendering follow-up is recorded separately as an exploratory evaluation, retaining checkpoint 1111 and all secondary checkpoints.

Rendering and model checks

Setting	Initial benchmark	Training-rendering follow-up
Native framebuffer	320×240	160×120
HUD	Off	On
Decals and particles	On	Off
Network input	128×72 RGB, CHW	128×72 RGB, CHW
Resize	Nearest neighbor	Nearest neighbor

Restoring these settings improved development success from 3/6, 4/6, and 4/6 to 6/6 for all three checkpoints. This tests the settings together; it does not isolate the contribution of HUD, resolution, or effects individually. The Basic WAD is byte-identical between the public training source and the installed ViZDoom scenario.

The APPO policy has 3,396,837 parameters, a CNN encoder, and a 512-unit LSTM. Evaluation strictly loads all 22 model-state tensors, including the saved input normalization statistics. Packed recurrent state contains 1,024 values for the hidden and cell states and resets at every episode. Inference preserves every parameter and normalization buffer.

Public checkpoints:

edbeeching/doom_basic_1111, revision fae3bfcd5ab5f98804fe4327a0be982776c8d7d3.
edbeeching/doom_basic_2222, revision 459b453cce290edeb946d1f4c755dbed974cf00b.
edbeeching/doom_basic_3333, revision 599f9aae7083798b207bfe05a9d0517c1da6de3c.

The runtime uses their original Sample Factory source, commit 9da68b57eecd73c3c884c1be2d938b46aa7a7f49 (2.0.0 source, using gym). Installed distribution metadata reports 2.1.1; the original source takes precedence on PYTHONPATH, and imported module hashes are recorded. Checkpoints load with weights_only=True and a scoped allowlist for legacy NumPy numeric metadata.

Verification and reproduction

The initial 432 episodes were independently replayed: 14,937 decisions, zero mismatches.
All 144 training-rendering follow-up episodes passed independent replay of 771 decisions, including framebuffer hashes, rewards, physical ticks, and final metrics.
Three APPO contract tests and thirteen existing Doom environment tests passed.
The existing Jev reference has 48 Basic episodes and 2,068 decisions; request receipts, action distributions, RNG decisions, and source hashes were audited.

The validation report records these checks. Raw episodes, model download manifests, checkpoints, and the pinned source archive are stored locally under runs/appo_basic_v1/. The existing Jev source remains data/unified_v2/jev_complete.jsonl.

With the pinned source and inference dependencies available on PYTHONPATH, reproduce the two evaluations using fresh output paths:

bash

python scripts/run_appo_basic_experiment.py \
  --model-dir runs/appo_basic_v1/models \
  --output-dir runs/appo_basic_repeat/episodes --workers 3

python scripts/diagnose_appo_doom_rendering.py \
  --model-dir runs/appo_basic_v1/models \
  --splits train,dev,calibration,test,ood \
  --output runs/appo_basic_repeat/native_rendering_all.json

summarize_appo_doom.py accepts all nine --run NAME=EPISODES inputs and an optional --rendering-report to regenerate the comparison. audit_appo_jev_reference.py rechecks the existing Jev reference without inference or API calls. Evaluation used CPU inference with one Torch thread per process and ViZDoom 1.3.0; no optimization or new Jev calls were performed.

The next Basic data collection can use these APPO actions as categorical targets for cross-entropy SFT. Its state export must use the actual rendering dimensions. Maze and Snake recipes are unchanged. No checkpoints were uploaded to Hugging Face.

快照未包含部分工程附件或配图；请通过上游链接或原始材料查看，收录范围见来源清单。