Abstract
The fleet queue ran at full settings on calber, columbia, nasser and the Mac. Its first 4 hours 10 minutes, counted from the logs it fetched back:
- 37,905 games of 15 minutes, which is 9,476 hours of play;
- 1.46 billion decisions (thinks), 10.4 billion seat-ticks and 381 million planned moves;
- 416 million training sample-steps on calber's GPU, in two net rounds.
The results so far:
- 144 weight trials, 6 of them past two errors. None has passed both held-out reads.
castle_inside8 leads: ten 55-seed censuses all at or above zero, and 2 of 5 held-out reads passed.- Both net rounds left bc2 the best.
I. Introduction
A. Intent
Keep the fleet's census running at the settings that can tell a weight from luck, then say how much it has computed.
B. The prompts
continue running census on the fleet.
also cafinate this mac
hows it going.
how many computations have been ran. over a million?
add castle_inside = 8 to the converge candidates. and add lab report about these counts.
Paul's choices:
- "Replace with full queue", which stopped the sample.
- "Wait for the 3 reads": castle_inside 8 joins the candidates only with the verdict of its held-out reads.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | The sample's 4-seed, 2-minute censuses cannot tell a weight from luck (CI-071 V) | Replace it with the full settings of docs/fleet-train.md | Leaving the sample running |
| D2 | stop-now kills jobs on three remote hosts; the permission check stopped it | Asked Paul; he chose to replace | Draining with stop-adding (slower; the sample's results matter little) |
| D3 | saskatchewan's load was about 20 of 24 on make converge | Left it out of the queue | Giving it fleet work (CLAUDE.md, memory) |
| D4 | Seeds must not overlap saskatchewan's run | Trials from 6,000,000, held out from 7,000,000; converge uses 2,000,000 and 3,000,000 | Sharing a seed range |
| D5 | "cafinate this mac" | caffeinate -dimsu in the background | Changing the Mac's power settings |
| D6 | "how many computations": a count must name its unit | Games, seat-ticks, thinks, moves and training sample-steps, each from the logs' own lines | Estimating from seat-ticks / 288,000 (net games have other sizes; the first reply's 32,000 games was that estimate) |
| D7 | A count must be repeatable | scripts/fleet-count.py reads only fetched results and events.jsonl | A one-off shell line |
| D8 | castle_inside 8 had passed 1 of 3 reads, not both of a pair | Asked Paul before adding it; he chose to wait | Adding it as a CANDIDATE, against LR 92's rule |
| D9 | The first report said three trials passed their re-checks | Corrected to Paul and recorded (LR 111 C) | Leaving it |
III. Methods
python3 ~/code/orrery/tools/fleet-queue.py build/queue/sample stop-now.build/queue/full-1/config.jsonis the sample's, with the planner's path changed. Itsplan.json: ``json {"commit": "8e5df8827fb680aa28ad3aad00e4fbb5308acfe0", "census_host": "columbia", "weights": {"seeds": 55, "secs": 900, "map": "all", "seed0": 6000000, "held_seeds": 400, "held_seed0": 7000000}, "net": {"rounds": [{"run": "rec2400", "lo": 9001260, "n": 1200}, {"run": "rec2400", "lo": 9002460, "n": 1200}], "train_args": "4 32768 3e-3", "census": {"seeds": 110, "seed0": 500000, "secs": 900, "map": "all"}}}``fleet-queue.py build/queue/full-1 ensure;caffeinate -dimsu &.- At 06:26:56:
python3 -I scripts/fleet-count.py build/queue/full-1. The held-out reads come fromplan/ledger.tsvand the job titles indone/.
IV. Results
| Kind | Games | Seat-ticks | Thinks | Moves |
|---|---|---|---|---|
| weight trials | 24,585 | 6,757,079,968 | 957,738,999 | 250,016,569 |
| held-out reads | 12,000 | 3,285,884,736 | 463,260,527 | 120,723,097 |
| net judging | 1,320 | 364,327,056 | 36,308,633 | 9,769,679 |
| all | 37,905 | 10,407,291,760 | 1,457,308,159 | 380,509,345 |
Net training:
- Round 1: 52,067,105 samples × 4 epochs, in 90.6 s.
- Round 2: 51,962,086 samples × 4 epochs, in 87.8 s.
Hours busy: calber 4.1, the Mac 4.1, columbia 3.8, nasser 3.3. No job failed.
| Trial | Weights | Held-out read 1 | Read 2 |
|---|---|---|---|
| w0011 | advance_band 0.675, alpha 0.513, answer_cap 2.51 | +0.56 ± 0.24 | −0.11 ± 0.24 |
| w0095 | castle_inside 8 | −0.02 ± 0.16 | +0.56 ± 0.21 |
| w0113 | base_r 8, beta 0.6, bridge_keep 360, castle_inside 8 | +0.19 ± 0.19 | +0.40 ± 0.19 |
| w0116 | column_gap 1 | −0.02 ± 0.17 | −0.20 ± 0.12 |
| w0117 | castle_inside 8 | running | +0.28 ± 0.21 |
| w0128 | castle_inside 8 | +0.80 ± 0.21 | running |
Net rounds, candidate less bc2 over 110 seeds:
- Round 1: −0.77 ± 0.36.
- Round 2: +0.18 ± 0.35.
Errors:
- Each count is a lower bound at its moment: running jobs are left out.
- The per-host games come from the ledger, which lags the logs by a census or two.
V. Discussion
- Yes, over a million, in every unit but whole games:
- 1.46 billion decisions;
- 10.4 billion seat-ticks;
- 416 million training sample-steps.
Whole games, at about 9,100 an hour, would reach a million in about 4.6 days.
- What the play bought so far:
- No weight trial yet passes the two-read rule. The defaults are near
the top of most curves, as saskatchewan's run found (journal 0103).
castle_inside8 is the one direction both runs point to.
- 19 trials changed no score at all. On these maps those weights are inert. They are cheap to drop from the screen, or worth a scenario that exercises them.
- Next:
- castle_inside 8's last two reads, then its line in saskatchewan's
candidates.txt;- the queue goes on.