jedarden bf80d84a6f notes(bf-413): document completion of AI Code Battle mechanics iteration

All backlog items completed:
- Combat kill scoring in engine (turn.go:274)
- Fitness formula blends win rate + kill rate (run.go:608)
- Flee thresholds reduced with outnumber logic
- CombatDeaths tracked through arena MatchOutcome
- Aggression derived from actual kill rate in behavior vector

This Genesis bead tracked the full mechanics iteration to make combat
economically necessary and reward aggression in the evolver.

2026-06-17 03:54:38 -04:00

3.6 KiB

Raw Blame History

Bead bf-413: Genesis - AI Code Battle Mechanics Iteration - COMPLETE

Status: ALL BACKLOG ITEMS COMPLETED

This genesis bead tracked ongoing work to improve ai-code-battle game mechanics so the evolutionary loop produces genuinely aggressive, combat-seeking bots.

Core Problem

Most bots flee before they enter attack range. Combat only happens between two non-fleeing bots (e.g. hunter+swarm). The evolver has no gradient toward aggression because fitness was pure win rate.

All Backlog Items Completed

✅ bf-z9m: Award score for combat kills in engine (turn.go) — P0

Commit: c1acd83 (2024-06-16)
Engine change: gs.Players[e.Owner].Score += gs.Config.KillScore (turn.go:272-275)
Combat kills now award points to the killer

✅ Reduce flee thresholds so bots only flee when outnumbered, not preemptively

Commit: d5515e0 (2024-06-17)
Changed flee threshold from AttackRadius2 + 4 to AttackRadius2 across all strategy bots
Added outnumber logic: only flee when nearbyAllies < nearbyEnemies
Modified files:
- bots/farmer/strategy.go: new shouldFlee() method with outnumber check
- bots/gatherer/strategy.go: shouldFlee() updated to check local counts
- bots/siege/strategy.go: shouldFlee() updated to check local counts
Bots now only flee when enemies are actually in attack range AND they are outnumbered

✅ Surface CombatDeaths through arena MatchOutcome into fitness formula

Commit: d42d1a5 (2024-06-17)
arena.go: CombatDeaths []int field in MatchOutcome
arena.go:204-221: Kill tracking and kill rate computation
turn.go:269-271: gs.CombatDeaths[e.Owner]++ tracks kills per player

✅ Derive aggression behavior vector from actual kill rate (not LLM self-report)

Commit: d5515e0 (2024-06-17)
cmd/acb-evolver/run.go:610-628: compute behavior vector from arena results
Formula: BehaviorVector[0] = min(killRate, 1.0) - aggression derived from actual kill data
Preserves existing economy value (BehaviorVector[1]) or defaults to 0.5 if not available
Enhanced logging to show derived aggression value in arena output

✅ Evolver fitness: blend win rate + kill rate (bf-59h)

Commit: d42d1a5 (2024-06-17)
run.go:601-608: fitness = 0.7*winRate + 0.3*killRate
cmd/acb-evolver/internal/prompt/builder.go:154-159: System prompt explains the fitness formula
Kill rate is now 30% of fitness, encouraging combat aggression

Design Principle Applied

Aggression must be economically necessary, not just rewarded. Reference: Google AI Challenge 2011 Ants — expansion forced encounters, combat was unavoidable.

In ACB, the flee logic previously created a false safe option that broke this loop. Now bots must fight when:

They are in attack range of enemies (no preemptive flee)
They are NOT outnumbered (flee is only for disadvantageous fights)

Recent Aggressive Seed Bots (commit `d42d1a5`)

Three new aggressive seed bots were added as generation-0:

Phalanx: Defensive cluster that moves as a group
Assassin: High-damage single-target eliminator
Opportunist: Waits for weakened enemies then attacks

These provide better initial aggression diversity for the evolver.

Verification Steps

To verify these changes produce more aggressive bots:

Run the evolver for several generations
Check that kill rates increase across generations
Check that bots no longer flee preemptively
Verify that behavior vector aggression values align with actual kill rates

bf-59h: Evolver fitness weight combat kill rate alongside win rate (COMPLETED)
bf-z9m: Award score for combat kills in engine (COMPLETED)

3.6 KiB Raw Blame History