Training needs a huge number of matches. If the real server, even sped up by the virtual clock, couldn't play enough of them in an hour, the project would need a separate, approximate simulator for bulk training and accept the gap between the two. This measured it directly: full 2v2 matches on the project's arenas, played through the same interface every agent uses, each run to its natural end: one side killed, or the game's own 46-minute time limit reached, which counts as a draw. Eight matches ran at once, their logins started 10 seconds apart. Eight is what was run, not a shown maximum for the machine.
Result#
The project's scripted baseline played one side in every match. Three opponents were tried:
| Opponent | Matches | Ended in a kill | Matches per hour |
|---|---|---|---|
| One that does nothing at all (earlier runs, one arena only) | 47 in each of two runs | all 47 | about 1,200 to 1,800 |
| One that plays legal moves at random | 47 | 47 | about 600 to 750 |
| The scripted baseline itself (the mirror) | 32 | 15 | about 65 to 72 |
In the mirror, 17 of the 32 matches ran out the 46-minute limit with neither side dead; the 15 that ended in a kill took 11 to 42 minutes of game time. The range in each rate is two ways of measuring it.
How fast the virtual clock ran. With eight matches at once, each match ran about 4 to 5 times faster than real time (the mirror's median about 5 times, the random opponent's about 4), against 10 to 25 times in earlier runs with fewer matches at once. That is the speed of eight matches at once with agents deciding inside the server's main loop, not the server's ceiling for one light match; the machine's CPU stayed around 13% of its 12 cores.
What limits the rate#
Two different things, depending on which comparison is read:
- Server work per tick did matter between the do-nothing opponent and the random one. Matches against the random opponent lasted only about a fifth longer in game time, but each tick of the server's main loop cost about 2.5 times as much, and that is most of why the rate fell by more than half. Each agent decides inside that loop, so seating real players costs server time directly. The machine itself stayed far from its CPU and memory limits.
- Match length explains the tenfold gap between the random opponent and the mirror. The mirror's ticks were no slower (if anything slightly faster), and its virtual clock ran at a similar speed; its matches simply run far longer, and over half end only at the time limit.
What it means#
The project decided against a separate simulator. The reason is which matches training plays first: as the project's training plan has it, training starts by imitating the scripted baseline and then plays against fixed opponents of several strengths. Of the opponents measured, only the two weakest gave matches at several hundred an hour or more. The one as strong as the baseline, the mirror, gave about 65 to 72 an hour, over half of them drawn at the time limit. That training's early matches will mostly run at the faster rates is an expectation, not a measurement; if they turn out as slow as the mirror, the question is open again.
Confidence#
Measured once for each opponent (twice for the do-nothing one), on the project's one machine, with its first scripted baseline, which is far below human play. The do-nothing opponent's runs used one arena; the other two used all four. A stronger baseline, a different machine, more matches at once or larger teams could change the numbers.