Home / One-Frame Breakout / 10x longer
MinAtar Breakout · follow-up
Breakout, 10x longer
Same game, players and protocol as One-Frame Breakout, but each life trains for 300 blocks (300,000 moments) instead of 30. 12 seeds per player. cadence-net 0.89.0.
The first run stopped while every curve was still climbing. So: what happens if you just keep going?
Bricks per test game
Median of 12 seeds, each point a 10-block window. Shaded: middle half of seeds for the position-term brain and the two-frame MLP.
What it showed
- Cadence keeps learning. All three Cadence brains more than doubled after block 30 and were still rising at block 300.
- One frame is a wall for a plain learner, not for Cadence. The one-frame MLP flattens near 6 bricks: it can't tell which way the ball is moving. Every Cadence brain beats it on 12 of 12 seeds, by about 7 bricks. A follow-up probe found why: Cadence wires in a working-memory copy of the last frame by default, so it has the ball's direction built in, not learned.
- Handing the MLP the second frame wins in the long run. At 30 blocks it was level with Cadence. By 300 it leads every Cadence brain (12/12 vs the position-term brain, 11/12 vs no-recall), about 6 bricks ahead. The probe says Cadence has nearly the same information (direction readable at 98% vs 99%), so the gap is in how it learns to use it.
- The 0.89.0 position term: null, confirmed. On these 12 seeds it looked like +3 bricks (7/12, p = 0.09), so we ran 24 more seeds. Pooled over 36 seeds it wins 17, median -1.0 bricks, p = 0.68. The lean was noise. Recall memory with or without it doesn't beat having no recall at all.
Numbers
| player | 30 blocks | 300 blocks [IQR] | agrees w/ coach |
|---|---|---|---|
| Cadence, recall + position term | 5.7 | 13.1 [11.8-13.7] | 94% |
| Cadence, recall, no position term | 5.6 | 10.4 [8.0-12.3] | 86% |
| Cadence, no recall memory | 7.0 | 13.3 [10.5-15.0] | 88% |
| Plain MLP, one frame | 5.1 | 6.0 [5.8-6.2] | 91% |
| Plain MLP, given two frames | 6.7 | 19.2 [17.7-20.4] | 90% |
Bricks per test game over the last 10 blocks, median over seeds. The 30-block column is the first run (separate lives, same seeds).
| paired, by seed | wins | median diff | Wilcoxon p |
|---|---|---|---|
| Cadence, recall + position term vs Cadence, recall, no position term | 7/12 | +3.0 | 0.0884 |
| Cadence, no recall memory vs Cadence, recall + position term | 6/12 | -0.6 | 1.0000 |
| Cadence, recall + position term vs Plain MLP, one frame | 12/12 | +6.9 | 0.0005 |
| Cadence, no recall memory vs Plain MLP, one frame | 12/12 | +7.2 | 0.0005 |
| Plain MLP, given two frames vs Cadence, recall + position term | 12/12 | +6.2 | 0.0005 |
| Plain MLP, given two frames vs Cadence, no recall memory | 11/12 | +5.9 | 0.0015 |
How it ran
One command on a desktop CPU, about 30 minutes per Cadence life and 3 per MLP life, 6 at a time: run.py --seeds 10-21 --arms pos32,pos0,norecall,mlp1,mlp2 --blocks 300 --procs 6 --out runs/long10x. Nothing else changed from the pre-declared protocol in run.py. Raw lives in runs/long10x.