Mario, level by level
One Cadence brain (cadence-net 0.91.1) plays the levels in order, learning the whole time. Two runs: a nearly blind brain (before), then a fresh brain with better senses (after). Updated 2026-10-31 06:48.
Learning fix: solo clears (updated 2026-10-31 06:50)
After the body-sense fix (LEARNING_FIX.md): one Cadence brain taught by a look-ahead coach's labels (DAgger), playing alone in every test. Start jitter = idle frames before the first button, which moves the enemies. SEL starts pick the checkpoint; TEST starts (32-41) and HELD (TEST plus starts 80-138, 50 in all) are never used for training or picking. Pre-registered bar: TEST 8/10. Stricter bar added later, because neighbouring starts behave alike: HELD 40/50.
Final numbers of the overnight run are on the showcase page. Best per level on 50 held-out starts: 1-1 50/50 (W1_F2ck110), 1-2 29/50, 1-3 45/50, 1-4 50/50 (each its own checkpoint). Best single brain for two levels (one_11_13, 10/30 20:55): 1-1 on 40 of 50 held-out starts and 1-3 on 42 of 50, the same weights for both. Its 1-3 TEST band alone is 6/10, so on the pre-registered 10-start bar it passes 1-1 (8/10) and not 1-3.
| level | brain | SEL | TEST | HELD | status |
|---|---|---|---|---|---|
| 1-1 | one_11_13 (st_13d ck_0013) | 15/24 | 8/10 | 40/50 | consistent on both bars, same brain as the 1-3 row below |
| 1-3 | one_11_13 (same brain) | 19/24 | 6/10 | 42/50 | consistent on the stricter bar, not on the 10-start TEST band |
| 1-1 | joint 1-1 + 1-3 brain (10/30 19:31) | 12/17 | 10/10 | 44/50 | consistent on both bars (88% of held-out starts) |
| 1-1 | 1-1_steady (10/30) | 10/10 | 10/10 | 38/50 | consistent on the pre-registered bar, just short of the stricter one |
| 1-1 | first clear brain (10/29) | 6/10 | 1/10 | not consistent: its "8 of 10" was on starts 1-12 only | |
| 1-2 | best (st_12cw ck_0011) | 1/17 | 6/10 | 8/50 | not consistent: clears sometimes, stuck for 3+ hours |
| 1-3 | joint 1-1 + 1-3 brain (same brain) | 17/17 | 3/10 | 16/50 | not consistent: the 17 practice starts it clears all share one frame phase (a decision is 4 frames); from the other three phases it stops at x 1,371. The picking starts now cover all phases. |
| 1-4 | 1-4 brain (st_14b ck_0010, 10/30 20:46) | 17/24 | 8/10 | 38/50 | consistent on the pre-registered bar, just short of the stricter one. Beats Bowser. The same checkpoint is only 23/50 on 1-1 and 0/50 on 1-3: not one brain for all levels yet. |
What still goes wrong: the brain is fragile. Its fit to the coach's labels barely moves (about 0.9, critical moments about 0.99), but its solo results swing between 10/10 and 2/10 from one 250-lesson step to the next. A level is a chain of dozens of must-get-right moments. Averaging weights, voting across checkpoints, a smaller step, a keep-only-if-better ratchet and longer teaching did not steady it. Saving and testing every step did: the consistent 1-1 brain is one checkpoint that was kept. On 1-2 that is not enough: the brain fits the coach's 1-2 labels only to about 0.85, and even a backprop network predicts only about 85% of them on screens it has not seen, critical moments included. Tested and ruled out as the cause: the coach (it clears 1-2 from every start tried), the eyes (a sharper view around Mario added nothing), the step size, more teaching, a wider or deeper brain.
First sight
How far it gets on a level before any practice on it: median x of 3 solo attempts.
blind brain new senses best baseline
What's happening
- One Cadence brain plays Super Mario Bros. in order, 1-1, 1-2, 1-3, ..., learning the whole time and never reset between levels. The test is first sight: 3 solo attempts on a level before any practice on it. If it were learning to play Mario rather than memorizing levels, first-sight distance would rise and beat simple baselines.
- Result so far: it does not. Neither brain beat the best baseline at first sight on any new level. Random buttons or an untrained brain went as far or further. No level was cleared solo. Practice helped on some levels (blind: 1-3 to 746, 2-2 to 1,496) and did nothing on others (1-2, 1-4).
- Before: the blind brain saw 26 x 32 grey blocks, 2 frames, nothing else. It began from last night's 1-1 training, so its 1-1 is not a real first sight. It forgot old levels fast: 1-1 went from 898 to about 300 after practising 1-2.
- After: a fresh brain with better senses. Eyes: 4 x 4 pixel blocks (4x the detail) plus how the last 3 frames changed, each frame normalized to its own background and contrast so overground, underground and castles look alike. Body sense: Mario's own speed, rising or falling, on the ground or not, facing, size, hurt and star timers, time left. From scratch it reached 594 on 1-1 in 40 minutes, then little else moved.
- Boundary: the body sense comes from the game's memory, but only about Mario himself. No enemy, block or pit positions are read; the world comes from the eyes.
- Practice on a level: it plays on its own and learns from the reward (going right, not dying), copies its own good stretches, and gets a few coached tries where a look-ahead planner writes down what it would have pressed. The planner never takes the controls. About a tenth of its play is an earlier level again. It moves on after a solo clear (flag in 2 of 3) or about 38 minutes.
First sight vs baselines
| level | blind | senses | run | jump | random | fresh | fresh+ |
|---|---|---|---|---|---|---|---|
| 1-1 | 594 > pre | 40 ≤ | 315 | 314 | 450 | 309 | 40 |
| 1-2 | 183 ≤ | 196 ≤ | 196 | 196 | 338 | 40 | 177 |
| 1-3 | 311 ≤ | 321 ≤ | 321 | 317 | 299 | 705 | 48 |
| 1-4 | 234 ≤ | 234 ≤ | 234 | 229 | 331 | 551 | 233 |
| 2-1 | 322 ≤ | 324 ≤ | 308 | 306 | 384 | 120 | 322 |
| 2-2 | 514 ≤ | - | 162 | 514 | 1,234 | 1,199 | 514 |
Median x, same 3 starts as first sight. run = hold run right; jump = hold run+jump; random = random buttons (10 attempts, fixed seeds); fresh = an untrained Cadence brain with the blind input; fresh+ = an untrained brain with the new senses. > beats the best baseline, ≤ does not. The blind brain's 1-1 is "pre": it began from last night's 1-1 training.
After practice
| level | blind best | cleared | senses best | cleared |
|---|---|---|---|---|
| 1-1 | 898 | no | 594 | no |
| 1-2 | 482 | no | 196 | no |
| 1-3 | 746 | no | 419 | no |
| 1-4 | 551 | no | 234 | no |
| 2-1 | 457 | no | 338 | stopped |
| 2-2 | 1,496 | stopped | - | - |
Best solo median x over the practice rounds (about 38 minutes per level). Cleared = flag in 2 of 3 solo attempts.
Videos
Learning-fix brains
Solo clips of the later brains, by file name. The Mario page has them with context and the brain roster.
Coaching used
New senses
| level | coached | labels | own | old play | min |
|---|---|---|---|---|---|
| 1-1 | 36 | 1,704 | 12,936 | 0% | 40 |
| 1-2 | 30 | 577 | 10,894 | 12% | 40 |
| 1-3 | 30 | 836 | 8,764 | 11% | 41 |
| 1-4 | 24 | 624 | 9,731 | 11% | 40 |
| 2-1 | 18 | 872 | 10,539 | 13% | 40 |
Blind
| level | coached | labels | own | old play | min |
|---|---|---|---|---|---|
| 1-1 | 54 | 3,372 | 18,632 | 0% | 40 |
| 1-2 | 45 | 2,282 | 16,473 | 11% | 38 |
| 1-3 | 70 | 2,495 | 20,617 | 13% | 39 |
| 1-4 | 56 | 1,858 | 22,696 | 12% | 38 |
| 2-1 | 30 | 2,182 | 18,882 | 11% | 40 |
| 2-2 | 12 | 1,013 | 14,470 | 11% | 35 |
coached: tries where the brain plays and a look-ahead planner writes down the button it would have pressed on each screen; it never takes the controls. labels: buttons it wrote. own: screens from the brain's own attempts that made progress. old play: share of self-play spent on an earlier level. min: practice minutes.
Retention
New senses: solo again after leaving 1-4 (18:47), median x of 3.
| level | now | its best | flags |
|---|---|---|---|
| 1-1 | 316 | 594 | 0/3 |
| 1-2 | 258 | 196 | 0/3 |
| 1-3 | 419 | 419 | 0/3 |
| 1-4 | 234 | 234 | 0/3 |
Blind: solo again after leaving 2-1 (15:26), median x of 3.
| level | now | its best | flags |
|---|---|---|---|
| 1-1 | 700 | 898 | 0/3 |
| 1-2 | 322 | 482 | 0/3 |
| 1-3 | 419 | 746 | 0/3 |
| 1-4 | 437 | 551 | 0/3 |
| 2-1 | 457 | 457 | 0/3 |