Lab report: the net in orrery

Abstract

orrery's Training view draws the net being trained: bc2, the clone of the scripted brain trained on 47 million moments of its play. It is a fully connected net, 143 senses in, two layers of 64 units, 47 scores out: 16,431 numbers. The pictures are explained here from measurements of its file, nets/bc2.bin, and of 200,000 recorded moments.

What the measurements show:

• 24 of the 143 senses are never non-zero, and their weights are still random.
• 2 of the 7 decisions (squad, signal) were never trained.
• Two senses carry the largest weights: how close a drone is (31.8), and the tank's speed (23.1).
• The first matrix's colour scale is set by its largest weight, while its median weight is 0.07. That is why the first matrix looks washed out.

The view needs, first of all: a colour scale that a few weights cannot wash out, progress over rounds, and what each round changed.

I. The Training view

Drawn by orrery at 1280 × 760 from the document scripts/fleet-view.py printed, after the run rec2400 had ended. From the top:

1. What is happening, in a sentence.
2. The round's three steps: record, train on calber's GPU, judge in the census on columbia. The current one lights up.
3. The fleet (left), a card a machine:
   • its role;
   • CPU load against cores (and calber's GPU);
   • what it is doing now in words, with the command doing it.
4. The three weight matrices (right), small: inputs across, outputs down. Click one to enlarge it.
5. The legend: red takes away from a unit, green adds.
6. The 47 scores by decision:

1. Agreement with the scripted brain on unseen matches, a bar per decision. The dark tick is bc1. The red tick is always choosing the commonest choice: a bar is only good well past its red tick.
2. The census against the official net map7: level.

II. Why the first matrix looks pale and striped

The 143 senses run across, the first layer's 64 units down. Each cell is one weight: green pushes that unit up, red down.

A. Almost everything is pale because one weight sets the scale.

• The scale runs to ±31.83, the matrix's largest weight.
• The median weight is 0.071, and 99% of weights are under 2.32.
• So a typical weight is drawn at about a twentieth of full strength.

B. The compass repeats every 11 columns (left two thirds).

• The first 88 senses are 8 sectors × 11 channels, so a channel's pattern shows 8 times across.
• Channel 0, a hostile tank that way, carries the most weight of any channel (mean |w| 0.43). The net attends most to where the enemy is.

C. The hints are a pale band (columns 124 to 132) because they were never on.

• All 9 were zero in every recorded moment. A weight on a sense that is always zero gets no training signal, so it keeps its random starting value (between ±1/√143 = ±0.084).
• Measured: mean |w| 0.041, what an untouched start gives.
• The same holds for:
  • the friend's-call channel, the last column of each group of 11;
  • the six contact senses describing a call.
• Scripted matches make no calls.

D. Column 92, the tank's speed, has the second-largest weight (23.1). Drive's commonest answer is stop, and speed tells stopped from moving.

E. Horizontal stripes are units that listen to many senses at once. Pale rows are units that barely listen.

F. Column 140, how close a drone is, has the largest weight in the net (31.8). It matches the gauntlet's G4: the clones dodge a strike drone 96 to 99% of the time, more often than the scripted brain (95.5%).

The orange cast is real: 59% of this matrix's weights are negative. That is common with ReLU units, which many weights keep quiet until the right senses are present.

III. Why the middle matrix looks like noise

The hidden units have no natural order. Shuffle the 64 units, and the matching rows and columns, and the net computes exactly the same thing. So a hidden-to-hidden matrix has no meaningful shape: only rows or columns that stand out mean anything.

It did learn: mean |w| 0.186, three times the 0.0625 of an untrained layer of this size, and its largest weight is 6.50.

IV. Why the last matrix has quiet bands at the bottom

G. Drive and lead are the strongest rows (mean |w| 0.24, 0.26).

H. Squad and signal (the bottom 11 rows) are faint, even noise: untrained. No label is recorded for them, so they keep their starting weights: largest 0.12, under the starting bound of ±0.125. What the net does with those two decisions today is random.

I. Vertical stripes are units many decisions use. Units 18, 7, 0, 60 and 33 have the strongest outgoing weights.

V. The interface today

Three views, chosen in the header or with Command-1 to 3. None of them changes anything yet.

• Lab: the real fleet, grove. calber, columbia, nasser and saskatchewan, from copal fleet state. They say "off bus" because they have not been enrolled: no fleet agent runs on them.
• Media: the SD cards and images the machines were installed from.
• Training:
  • Refreshes every 5 s from one document (about 140 KB), which

  scripts/fleet-view.py prints in about 1.5 s: each machine over ssh, the newest recording run, the loop's state, the best net and its results.

  • Text is drawn a pixel bolder for legibility.
  • Clicking a matrix enlarges it, with labelled axes, a legend and the

  weight under the pointer.

VI. What it needs to show how training goes

The view is a snapshot of one net. Analysing training is seeing change. In order of value:

1. A colour scale a few weights cannot wash out: 99th-percentile scaling, clipped cells marked, a histogram of weight sizes.
2. Progress over rounds: loss per epoch, agreement per decision and the census with error bars, as lines, with the official net as a reference.
3. What each round changed: candidate minus best, per matrix.
4. Which senses each decision listens to: a 7 × 7 map of decisions by sense group.
5. Where each decision goes wrong: a confusion table per decision (the drive head's stop habit would be one bright column).
6. Which inputs are alive: a strip under the first matrix of how often each sense is non-zero, with dead senses and untrained decisions marked.
7. Healthy units: units never active or always active on held-out moments.
8. The 110 scripted-brain weights as the trainer tunes them: a curve per weight, the trials queued, the candidates waiting.
9. The queue and its controls: jobs by kind, machines' use over time, "stop adding" and "collect results and stop now" in operator posture.
10. Names on hover: "sector 3, hostile tank", not just "compass".
11. Two nets side by side, and a choice of round.
12. Fast and slow refreshes apart: machines every few seconds, the net when a round ends.

VII. Method

The figures here come from three places:

• The weights: read from nets/bc2.bin (Net::save's TCNET1 form: each layer's weights output-major, then its biases).
• The senses: mean, spread and share non-zero, from 200,000 samples of shard 9000060 of the run rec1200. The layout is from senses.rs (SECTORS × CHANNELS, BODY, NAV, CONTACTS, HINTS, SHELL, AIR).
• The starting bounds: Burn's default for a linear layer, ±1/√inputs.

The pictures were drawn by orrery's --frame at ccb3ae2, with ORRERY_FRAME_ZOOM=N for the enlargements.