Abstract
A review, no run. It does three things. It lists the methods the census gained in LR 78 to 86: the sequential census, families as a guard, parts taken apart, the squad and spending meters, rungs.py and traits.py, rewards in the score, and worktrees on the machine. It sets the gains beside how they were found. Continuous tuning by SPSA read flat (LR 74: -0.00 +/- 0.14). Turtle, a hand-set style, read +1.40. The cowboy, new behaviour, read +11.62. And it plans the census as a learning loop: the paired score as the objective, fresh seeds as batches, 400 held-out seeds as the test set, families as constraints, the NOTES rule as preregistration. The steps, in order: make dead weights act and screen every weight's slope; evaluate against a league of fixed opponents; run MAP-Elites over the personalities with traits.py's signature as the descriptor; batch search on the 24 threads; weights by map features.
Index terms: methods, brain census, machine learning, SPSA, quality-diversity, MAP-Elites, league training.
I. Introduction
A. Intent
To commit LR 85 and 86 on main, and write the report Paul asked for.
B. The prompts
looking good. the census just finished. can we get a visual playtest. I want to see these battles in a 16 bot map. show 2 of the screen.
merge. only on main. i think we are ready. lab report about these new methods and looking forward into running census and using this like machine learning.
II. Decision Log
| # | Cause | Decision | Not taken, and why |
|---|---|---|---|
| D1 | "merge. only on main" | LR 85 and 86 committed and pushed on main (bd08f40); no branch | A branch and a merge: the project works on main |
| D2 | "a visual playtest ... 16 bot ... 2 of the screen" | CLIENTS=2 PLAYERS=16 TEAMS=2 scripts/quad.sh on the machine's screen, from the worktree with the new brain | The machine's main tree: it was behind the cowboy then |
| D3 | The report looks forward | LR 87 a review and a plan, no run | A first learning run in the same step: one step at a time |
| D4 | LR 74 flat, LR 86 large | The plan puts new behaviour and dead weights before tuning | Start with a bigger optimiser: it would polish a ridge |
| D5 | Shelling pillboxes mostly paid by its reward | The old score reported beside the rewarded one, as a guard on every learner | Drop the rewards: Paul asked for them, and they are visible now |
| D6 | The defaults move | A league of fixed opponents | Climb against today's defaults alone: a cycle would go unseen |
III. Methods
The review read LR 24 to 86, the ledger (ledger.rs), scripts/ spsa.sh (LR 40) and the census timings on the machine (a 110-seed census about 7 to 9 minutes, 400 seeds 20 to 30, every map, 24 jobs).
IV. Results
In LR 87: the methods table; the gains table; the mapping of learning to the census; the plan and its guards.
V. Discussion
The census is already most of a learning system: it has an objective, fresh data each step, a held-out test, constraints and preregistration. What it lacks is a search that proposes behaviour rather than numbers, and opponents that stay put. The personalities Paul asked for are the bridge. A quality-diversity search keeps one of each, tuned, and plain to see by construction.