Weighted curriculum
0.00sLoading policy…
New: guard refinement · three variants, five seeds & best checkpoints →
BEHAVIOR LAB / 03 — STARTING FROM ZERO
TEN MILLION.
EACH.
Weighted curriculum versus potential shaping.
Two fresh guards. The same ratings, rules and movement options.
—TRAINING REPS COMPLETED
ZEROINHERITED TRAINING
10KHELD-OUT REPS / TEST
FIXEDRATINGS & PHYSICS
WEIGHTED CURRICULUMStarting…
POTENTIAL SHAPINGStarting…
Loading experiment…
01 / HELD-OUT COMPARISON
Testing a longer training budget.
Final evaluations appear after both methods reach exactly 10 million training reps. The test set is separate from training and checkpoint monitoring.
| Guard policy | G win rate ↑ | Lane coverage ↑ | Lane error ↓ | Wrong-way motion ↓ |
|---|---|---|---|---|
| Final evaluation pending. | ||||
One training run per method. A large test set measures these two saved policies precisely; it does not measure variation across independent training runs. No reliable method-wide winner can be established from this run alone.
Lane coverage: G is between DT and QB and within 0.8 yd of their connecting line. Wrong-way motion: free-movement time traveling laterally away from that line while misaligned. These are positioning proxies, not NFL technique scores.
02 / SAME SCENARIO, DIFFERENT TRAINING
Curriculum. Shaping. Side by side.
Choose a saved checkpoint to watch development from zero. Both guards face the same setup and DT plan. Each uses its own learned opening and movement decisions. Replays never update their policies.
Potential shaping
0.00sLoading policy…
GUARDDTQBDashed line: direct rush laneTrails: player movement
Loading matched pair…
These study bundles include their training method. Keep them with this viewer and the reproduction tools. The main lab remains on its existing players and outcome-only training.
03 / CHECKPOINTS
What changes with more practice?
Both policies face the same 2,000 development scenarios at every checkpoint. These scores are separate from the final held-out tests. Early curriculum checkpoints cover only the stages practiced so far.
Learning curves will appear with the final report.
| Reps / guard | Curriculum wins | Shaping wins | Curriculum coverage | Shaping coverage |
|---|
Curriculum changes at 2M and 5M reps. Scores at those checkpoints are measured just before the next stage starts. A checkpoint is not promoted just because it has a high development score.
04 / TRANSFER & STEERING CHECK
Protection beyond the practice mix.
Each final test contains 10,000 matched scenarios per policy. “Authored steering” keeps each policy’s learned opening tactic but disables its learned movement adjustments. The million-rep DT remains frozen.
| G win rate | Untrained | Curriculum · 10M | Shaping · 10M |
|---|---|---|---|
| Final evaluation pending. | |||
05 / EXPERIMENT RULES
A fresh start. A fixed comparison.
Weighted curriculum
0–2M: head-on bull rush.
2–5M: bull, left arc and right arc.
5–10M: 50% normal rush mix, 30% lateral/feint rushes, 20% recovery starts with G offset ±1.7 yd.
Position learning receives the existing win/loss reward. No shaping bonus is added.
Potential shaping
Uses the full normal rush mix throughout. Position reward adds 0.25 × [0.94 Φ(next) − Φ(current)], with lane alignment 50%, pocket cushion 30%, and movement control 20%. Terminal Φ is zero.
No curriculum is applied. Both methods keep the same outcome-only opening-tactic learner.
Initialization, validation & limitations
Both learners start with zero reps, empty movement tables, and zero-valued opening tactics. Neither inherits the earlier study nor the million-rep player policies. DT does not train in either arm. Guard ratings remain strength 89, speed 72, technique 70, 315 lb; DT 74, 87, 70, 280 lb.
Physics, observation bins, movement options, 0.3s movement decision cadence, learning rates and exploration floors stay unchanged. No parameters are tuned after inspecting these runs. Longer training does not guarantee monotonic improvement.
The headless trainer skips visual logs/trails and unnecessary coordinate serialization. Each method matched the original engine over 10,000 rep-by-rep trajectories and final learned weights before the full run began. Final policy evaluations and visible replays are also checked against the original browser engine.
Training seeds are unique and disjoint from evaluation seeds. Checkpoints and resumable saves are retained. One seed per method cannot establish how reliably a method improves other independently trained players.
Download experiment manifest ↓ · Earlier three-seed screening study →