Skip to content

AI Training

Jump to bottom
Dudiebug edited this page Apr 30, 2026 · 2 revisions

AI Training

TerminatorPlus can train movement-controller brains in the JVM. The neural network controls movement only; CombatDirector remains the sole owner of combat actions and item decisions.

Start Training

/ai reinforcement <population-size> <name> [skin] [mode-or-options] [round-minutes]

Examples:

/ai reinforcement 120 TrainBot Steve
/ai reinforcement 120 TrainBot Steve family=mace:mix=mace_curriculum 5
/ai reinforcement 120 TrainBot Steve movement:family=mace:mix=mace_curriculum 5
/ai reinforcement 80 TrainBot Steve legacy

If [round-minutes] is omitted, the command uses ai.training.max-round-minutes from config.yml. The default is 1; set it to 0 for unlimited rounds. Optional command arguments are positional, so pass movement or an options string in the mode slot when you want to override only the minute cap:

/ai reinforcement 120 TrainBot Steve movement 5

Modes:

  • Empty mode: train the movement brain bank.
  • movement, movement_controller, or movement-controller: train the movement brain bank.
  • legacy: original full-replacement neural-network training.

Automatic Loadout Assignment

Movement training samples named loadouts automatically from ai.training.loadout-mix. No separate /bot loadoutmix step is needed.

The default movement_balanced mix is:

sword: 12
axe: 12
smp: 12
pot: 8
mace: 10
spear: 8
trident: 8
windcharge: 6
skydiver: 5
hybrid: 6
vanilla: 5
pvp: 3
crystalpvp: 3
anchorbomb: 2

pvp, crystalpvp, and anchorbomb total 8%, which gives explosive/survival movement enough exposure without letting those kits dominate every generation.

Curriculum Families

Set ai.training.curriculum-family or pass family=<name> in the training options argument to train one specialist family:

  • melee
  • mace
  • trident_ranged
  • spear_melee
  • mobility
  • explosive_survival
  • projectile_ranged

Curriculum mixes are configured under ai.training.loadout-mixes, including melee_curriculum, mace_curriculum, trident_curriculum, mobility_curriculum, and explosive_survival_curriculum.

Mixed training seeds each candidate from its assigned loadout family, ranks results by the movement family that actually produced route samples, and updates every eligible specialist brain represented in the round. A family brain is not saved from unrelated survival or aggregate fitness. Curriculum mode forces all candidates to update the configured family brain.

Reward Profiles

Movement reward scoring is family-specific:

  • general_fallback: balanced damage delta, survival, range control, low fallback rate, and low route thrash.
  • melee: melee threat range, legal crit/sprint-hit setup, and hit conversion.
  • mace: launch/airborne/smash phase conversion, tracking, and self-damage avoidance.
  • trident_ranged: charge completion, line of sight, throw range, and hit rate.
  • spear_melee: close trident pressure and stable melee spacing.
  • mobility: gap close, escape, vertical setup, and route handoff success.
  • explosive_survival: target explosive damage, safe blast spacing, escape, and low self-damage.
  • projectile_ranged: line-of-sight control, lateral strafing, and projectile hit/interrupt success.

Rollout metrics are saved with the trained brain and surfaced by evaluation reports.

Commands

/ai stop
/ai brain status
/ai brain load
/ai brain save [bot-name]
/ai brain reset
/ai movement <amount> <name> [skin] [loc]
/ai evaluate [variant] [scenario] [seeds]

/ai movement spawns fighting bots using the loaded movement bank. They are not training bots.

/ai evaluate exports a report-only seed/scenario matrix and route/fallback summary under ai/movement/evaluations/. Live win/damage metrics require an arena run.

See Movement Brain Bank and Brain Persistence for the bank layout and file formats.

Clone this wiki locally