Flee or fight? An A/B experiment on 100 players
Fleeing disappoints: of 52 flights, 22 succeeded and 28 ended in death. Rather than guess, the experiment is on: half of the autonomous players still flee when a fight goes badly, the other half fights to the end with its emergency moves. Death reports record the group; the automatic analysis compares. Also, “on next swing” attacks are no longer cancelled by their own re-cast, and isolated targets are strongly preferred.
- Commit
- 82e5299
- Build
- Build success (2 min) · Target worldserver
Previous window (10 min)
- 44 deaths (45 before): no clear progress.
- Emergency moves well used: Evasion 16 times, Hammer of Justice 13, Gouge 11.
- Flights: 52, of which 22 “successful”, 1 “caught”, and 28 deaths while fleeing.
- Adds: 26 of 37 come from other species (nearby monsters, patrols), 11 from kin called for help.
- Reason 40 (
SPELL_FAILED_INTERRUPTED) on Heroic Strike and Raptor Strike: the brain re-cast these “on next swing” attacks at each decision, cancelling the previous one.
The experiment
FreeWill.FleeAbTest = 1: characters with an even GUID keep fleeing, those with an odd GUID never do (about 50 players per group). The flee_group field of death reports lets the analysis compare both groups' mortality over 30 minutes; the winning rule will be applied to everyone.
Other fixes
- An “on next swing” attack already queued is no longer re-cast.
- Isolated targets first: each hostile neighbour within 18 m of a target weighs heavily in its score; an accompanied target is chosen only when nothing better exists.