Abstract
The fleet's two machines now read the same single-decision tests for navigation and takeover. A scenario's terrain may be a wall, and six gauntlet files put a maze shaped S, 5 or @ between a tank and a pillbox it must take, or a pillbox and a base to take or to build. The rules, on 200 runs each, pass the S 99.5%, the 5 66.5%, the @ 100%, take a pillbox 99.5%, take a defended base 3.5% and build a base 88.0%; their floors are today's success less two errors. Crucible plays the same files' words on the GPU.
I. Intent
Paul: "Specifically the goal is to naviate a simple maze, Like a maze of S shaped and 5 Shaped, and @ (ampersat at-sign) shaped mazes on test gauntlet maps of player vs environment, where we add takeover takss like Pillbox capture, Base capture, Building a base capture. A series of rewards need to be placed into the GPU and CPU such they can simulate the same score, or at least correlate their fitness." And: "Code the implimentation of more tests, coded in Cuda, for GPU and CPU." Then, asked what to do with the work: "Commit and push both."
II. Decision log
| Decision | Cause |
|---|---|
| A wall as a terrain kind, nothing else in the loader | the gauntlet's files already lay rectangles of terrain; a maze is rectangles of wall |
| The mazes take a pillbox at their end | the net's senses point at the nearest pillbox, so a GPU net can be asked the same task; the rules navigate by the map |
pillbox_owned_by as the pass | treadcore's word for a takeover; crucible reads it as its own capture |
| Floors at today's success less two errors at 200 runs, G25 at 0.95 | LR 141's rule; a 100% share has no error, and a floor of 1.00 fails a promotion on one run |
| The same coordinates in both gauntlets | so a share on the card and a share on the CPU are shares of the same maze |
III. Method
GAUNTLET=1 GAUNTLET_BRAINS=scripted GAUNTLET_RUNS=200 cargo test -p treadcore-brain --release --lib -- --ignored learn::harness::train, on the Mac, seeds 9,500,000 onward as every gauntlet read; the brain's unit tests, cargo test -p treadcore-brain --release --lib.
IV. Results
| File | The rules, 200 runs | Floor |
|---|---|---|
| G23 the S maze | 0.995 | 0.95 |
| G24 the 5 maze | 0.665 | 0.55 |
| G25 the @ maze | 1.000 | 0.95 |
| G26 take a pillbox | 0.995 | 0.95 |
| G27 take the base | 0.035 | 0.00 |
| G28 build a base | 0.880 | 0.80 |
The brain's 168 tests pass; the whole gauntlet, the six included, reads in 1.8 s at 20 runs.
V. Discussion
The 5 is the hard maze for the rules: its bowl turns back on itself, and a third of the runs end short of the pillbox. A defended base is nearly never taken: the enemy's tank and pillbox together outgun one tank, and G27's floor is zero until the rules learn it. On the card the GPU's nets passed none of the mazes on their first day, and one of them built a base 56 times in 256 after two segments of training on the tasks.
VI. What is not built
The mazes are open fields with walls, not the fleet's maps; the takeover's squad and carrying are the game's own and not tested here.
VII. Paul's decisions
Commit and push (6 November).