The project on one page

The game

Treadcore is an overhead tank game for up to sixteen players on a 128 by 128 tile island, one tile 25 metres. It descends from a 1980s tank game and adds what the war of the last few years handed the tank: the gun fires indirect at a point a recon drone is watching, interceptors hold a ring round the hull and ram what flies in, strike and longshot drones carry the fight past the horizon, and jammers blind everyone's link alike. Each tank carries a squad of four that captures pillboxes, builds walls, roads, bridges and mines, fells trees for material, and can lift a pillbox and carry it home. Depots resupply. Alliances form and break under a three-way rule. The windows find each other on a LAN or across the Internet and elect a host; nobody runs a server. It is Rust, with Bevy for the window; the simulation is deterministic and headless, and everything above it drives it through two calls, step and observe. The player's guide is How to Play; the rules and their reasons are DESIGN.md.

The two brains

Every seat a person does not take is a bot, and a bot is one of two kinds of mind, or a mix.

The programmed brain (the Commander, in treadcore-brain) is a ladder of goals a person can also set by hand: defend the base, fortify, hunt pillboxes, expand, guard. Under it sit about 135 named numbers, the weights: some are worths that lean a choice, some are shares between 0 and 1, some are distances and times with a meaning in the world, some are whole-number choices that pick one of several rewritten versions of a function, and a few are on-off switches. A team shares one board (influence, the front, posture, jobs) and plays roles: a leader, a guard, rangers, wingmen, and on a person's team a builder. Eight named styles (Fleet, Raider, Castle, Expander, Timid, Overrider, Scholar, Duelist) are whole brains anybody can pick by name and put on a side. The month's record of building it is the brain's lab report, 126 numbered sections.

The learned brain is small on purpose. A tank's senses are 133 numbers (a compass of sectors, the body, the instruments, a periscope's contacts, hints); a net of two hidden layers of 64 reads them and chooses. Beside it sits the transform: a few hundred numbers that read the senses, the slow waves of a tank's temperament and the rules' own answer, and may move that answer when confident. All zeros is the brain as it was, so training starts exactly where the rules stand. A GPU arena, crucible, breeds a third kind: duelling champions on islands of a population, whose best plays the fight inside a rules tank on this simulator. The plan for the learned tank is AI-BRAINZ-PLAN.md; how it is trained, and the parity between CPU and GPU, is LEARNING-DESIGN.md.

The instrument

Orrery began as a fleet console with no dependencies: a window drawn by hand on a Mac and on Wayland, SSH, SFTP, TLS and RDP written from the standards, so that a locked-down node on a LAN that never reaches the Internet could be reached, read and commanded. In the last week of October it grew the second half of its job: it holds the orchestration of training. A keeper runs a queue; each machine is a lane; a planner decides every minute what the lanes should do. The work is a census: seeded matches of a trial brain against the shipped one, pooled by inverse variance, and judged against bars written before the run. A trial that passes its reads and two held-out reads becomes a generation on a local line; the line is proved against the shipped game at 800 seeds, hourly, and goes to main past z 3. Every number of that process a person might want to move is a value on a screen; everything a person might need to do is a button; what the machines found is drawn, with hover, on the Results chapter's tabs. The Double Report is the guide to every screen; The Whole Line is the process end to end.

The definition of the design

Paul set the frame on 2 November in six parts, and this book keeps to it:

1. The design: the game's rules, the brain's shape, the process. Written, mostly; chapter 2 says where.
2. The implementation: the game, the brains, the console. Built and shipped (0.2.0).
3. Training the programmed brain's weights: the fleet's censuses, the line, the proof. Running; a week of it is this book's subject.
4. Training the transform into a learned player: built, with a first trained transform in a slot; not yet promoted, and the width test said wider is not yet worth it.
5. Combining the two: the Overrider style (rules under a transform) and the Duelist (rules with a champion in the fight) exist; the net cannot yet build or crew.
6. Orrery orchestrating the whole: dependencies, the order of seeds, promotion, values. Built; its faults were the week's main lesson.

Where it stands, in one table

Chapter 2 rewrites the goals; chapter 5 is the how-to for the line; the plan begins at chapter 9.