"Agent" on these pages means a game agent: the trained player. The AI agents that build and run the project are covered under autonomy.
The shape of it#
real game server ──► perception layer ──► observation ──► agent
▲ (what a player could │
│ know, when they'd know it) │
└──────────── agent interface ◄──── action ◄─────────┘Every agent, however it works, is seated on a bot character on the real server. It gets observations only from the perception layer and acts only through the agent interface. It never holds a game connection of its own, so there is no way around the limits.
What is fixed and what isn't#
| Fixed, so results stay comparable | Replaceable as methods improve |
|---|---|
| The real game server | Agents and their models |
| The perception layer and its principles | Training methods |
| The agent interface | Simulators |
| The loadout and comp format | Compute |
| How strength is measured |
The plan for agents#
- A scripted baseline for each starting team: hand-written play that is decent and predictable. It is the first opponent and the fixed yardstick.
- Learning takes over wherever it proves stronger than the script, measured by rating on the real server.
- A gradient: earlier agent versions fill the ratings below the top, so a human of any skill finds a fair opponent.
Pages in this section#
| Page | What it covers |
|---|---|
| Human parity | What agents may perceive and how fast they may react: the parity principles, and the one layer that enforces them. |
| The harness | The machinery agents play through: bots created per match, one interface designed for in-process and out-of-process play, the match recorder and the virtual clock. |
| Measurement | How strength is judged: win rate on the real server, ratings with confidence intervals, fixed anchors, and how many matches it takes to tell two agents apart. |
| Server speed | The first measurement: how long the game server takes per step during an arena match, which sets how much faster than real time training matches could run. |
| Matches per hour | Whether a bulk training simulator is needed, measured directly: how many full arena matches the project's one machine plays each hour, each run to its natural end (a kill or the game's time limit). |