Document human 3 vs AI behavior findings

2026-07-28 21:35:42 +00:00 · 2026-07-14 07:00:36 -07:00
parent 15d7d30ccf
commit aaa3457f69
1 changed files with 227 additions and 0 deletions
@@ -0,0 +1,227 @@
+# Human 3 vs AI 4+ Behavior Analysis
+
+Source save: `game_1f949539969b79f6.zip`
+
+Downloaded/analyzed: July 14, 2026
+
+Game span in structured decisions: August 371 through September 391.
+
+Human player: faction 3.
+
+Comparison groups:
+
+- `Human 3`: the human player.
+- `AI 4-6`: surviving equivalent-start AI factions in the latest save.
+- `AI 7-9`: equivalent-start AI factions that were reduced or eliminated earlier.
+- `AI 4+`: all equivalent-start AI factions together.
+
+Factions 1 and 2 are intentionally not the main comparison group because their starts are structurally different.
+
+## Executive Summary
+
+Human 3 plays materially differently from AI factions 4+ in ways that plausibly explain better survival and growth:
+
+- Human 3 trains far more often, especially in hostile-border provinces.
+- Human 3 rests almost never.
+- Human 3 recruits and divines more, creating a much larger hero base.
+- Human 3 uses Feast more and HeroGift less than the surviving AI.
+- The surviving AI factions spend more turns on Improve and Rest, especially when small.
+
+The clearest next target is Rest: AI 4-6 still rests far more than the human player, especially at 1-2 provinces.
+
+## Overall PlayerCommands
+
+These counts exclude battle aftermath and other non-player-command phases.
+
+| Command | Human 3 | AI 4-6 | AI 4+ |
+| --- | ---: | ---: | ---: |
+| Improve | 152 / 520 = 29.2% | 393 / 993 = 39.6% | 546 / 1449 = 37.7% |
+| Train | 61 / 520 = 11.7% | 47 / 993 = 4.7% | 49 / 1449 = 3.4% |
+| Rest | 4 / 520 = 0.8% | 93 / 993 = 9.4% | 169 / 1449 = 11.7% |
+| Feast | 45 / 520 = 8.7% | 59 / 993 = 5.9% | 77 / 1449 = 5.3% |
+| HeroGift | 8 / 520 = 1.5% | 35 / 993 = 3.5% | 41 / 1449 = 2.8% |
+| March | 35 / 520 = 6.7% | 73 / 993 = 7.4% | 188 / 1449 = 13.0% |
+| RecruitHeroes | 33 / 520 = 6.3% | 23 / 993 = 2.3% | 26 / 1449 = 1.8% |
+| Divine | 19 / 520 = 3.7% | 13 / 993 = 1.3% | 18 / 1449 = 1.2% |
+
+The all-AI March rate is inflated by AI 7-9. Surviving AI 4-6 marched at roughly the same overall rate as Human 3.
+
+## Size-Normalized Differences
+
+The largest differences remain visible when comparing similar faction sizes.
+
+### 3-4 Provinces
+
+| Command | Human 3 | AI 4-6 |
+| --- | ---: | ---: |
+| Train | 25 / 110 = 22.7% | 5 / 220 = 2.3% |
+| Rest | 2 / 110 = 1.8% | 1 / 220 = 0.5% |
+| Feast | 5 / 110 = 4.5% | 31 / 220 = 14.1% |
+| RecruitHeroes | 4 / 110 = 3.6% | 9 / 220 = 4.1% |
+| March | 3 / 110 = 2.7% | 20 / 220 = 9.1% |
+
+At 3-4 provinces, Human 3 is much more training-heavy. AI 4-6 uses Feast and Alms more in this bucket.
+
+### 5-6 Provinces
+
+| Command | Human 3 | AI 4-6 |
+| --- | ---: | ---: |
+| Train | 19 / 139 = 13.7% | 3 / 248 = 1.2% |
+| Rest | 0 / 139 = 0.0% | 4 / 248 = 1.6% |
+| Feast | 15 / 139 = 10.8% | 7 / 248 = 2.8% |
+| RecruitHeroes | 9 / 139 = 6.5% | 6 / 248 = 2.4% |
+| March | 14 / 139 = 10.1% | 37 / 248 = 14.9% |
+
+At 5-6 provinces, Human 3 keeps investing in Train, Feast, and recruitment. AI 4-6 still prefers Improve and March more often.
+
+## Hostile-Border Provinces
+
+Hostile-border provinces are where strategic readiness matters most.
+
+| Command | Human 3 | AI 4-6 |
+| --- | ---: | ---: |
+| Improve | 102 / 348 = 29.3% | 127 / 428 = 29.7% |
+| Train | 58 / 348 = 16.7% | 16 / 428 = 3.7% |
+| Rest | 2 / 348 = 0.6% | 36 / 428 = 8.4% |
+| Feast | 36 / 348 = 10.3% | 14 / 428 = 3.3% |
+| HeroGift | 6 / 348 = 1.7% | 25 / 428 = 5.8% |
+| ArmTroops | 12 / 348 = 3.4% | 25 / 428 = 5.8% |
+| Diplomacy | 4 / 348 = 1.1% | 32 / 428 = 7.5% |
+
+Improve rates are nearly identical on hostile borders. The big behavioral split is what happens around that Improve baseline:
+
+- Human 3 adds much more Train and Feast.
+- AI 4-6 adds much more Rest, HeroGift, and Diplomacy.
+
+## Command Choice When Available
+
+Choice rate when a command was available in the acting province is a better signal of priority than raw counts.
+
+### All PlayerCommands
+
+| Command available | Human 3 chose | AI 4-6 chose |
+| --- | ---: | ---: |
+| Improve | 152 / 401 = 37.9% | 393 / 777 = 50.6% |
+| Train | 61 / 382 = 16.0% | 47 / 718 = 6.5% |
+| Rest | 4 / 404 = 1.0% | 93 / 872 = 10.7% |
+| Feast | 45 / 395 = 11.4% | 59 / 803 = 7.3% |
+| RecruitHeroes | 33 / 34 = 97.1% | 23 / 23 = 100.0% |
+| Divine | 19 / 28 = 67.9% | 13 / 28 = 46.4% |
+
+The AI does recruit when recruitment is available. The recruitment gap is upstream: Human 3 appears to generate or discover more recruitment opportunities, helped by more Divine usage.
+
+### Hostile Border, 3-6 Provinces
+
+This is probably the most useful apples-to-apples slice.
+
+| Command available | Human 3 chose | AI 4-6 chose |
+| --- | ---: | ---: |
+| Train | 42 / 156 = 26.9% | 6 / 155 = 3.9% |
+| Rest | 2 / 156 = 1.3% | 1 / 162 = 0.6% |
+| Feast | 15 / 154 = 9.7% | 9 / 158 = 5.7% |
+| HeroGift | 4 / 154 = 2.6% | 17 / 159 = 10.7% |
+| March | 8 / 156 = 5.1% | 17 / 162 = 10.5% |
+| ArmTroops | 7 / 13 = 53.8% | 22 / 37 = 59.5% |
+| Diplomacy | 3 / 153 = 2.0% | 14 / 151 = 9.3% |
+
+The Rest gap is not large in this exact slice, which suggests the recent Rest changes are helping once AI factions reach mid-size. The Training gap remains very large.
+
+## AI Decision Reasons
+
+Top AI 4-6 reasons:
+
+| Count | Reason |
+| ---: | --- |
+| 220 | `chosen neutral neighbors: Improve b/c nothing better: ImproveCommandSelector.commandSelection` |
+| 121 | `traveling` |
+| 63 | `chosen no neutral neighbors: Improve b/c nothing better: ImproveCommandSelector.commandSelection` |
+| 49 | `chosen neutral neighbors: rest` |
+| 46 | `maybeAlmsCommandForProvince` |
+| 45 | `Get support up in leader province` |
+| 44 | `chosenUniversalCommand: rest` |
+| 30 | `offer truce with hostile neighbor to buy development time` |
+| 29 | `chosen neutral neighbors: train b/c nothing better: chosenTrainCommand` |
+| 25 | `chosenUniversalCommand: rest: feast instead of rest` |
+
+AI Rest reasons:
+
+| Count | Reason |
+| ---: | --- |
+| 120 | `chosenUniversalCommand: rest` |
+| 49 | `chosen neutral neighbors: rest` |
+
+AI Train reasons:
+
+| Count | Reason |
+| ---: | --- |
+| 29 | `chosen neutral neighbors: train b/c nothing better: chosenTrainCommand` |
+| 14 | `fulfill UpgradeBattalionQuest (train progress)` |
+| 5 | `prepare attack with hostile neighbors: train troops in preparation for attack: chosenTrainCommand` |
+| 1 | `chosen no neutral neighbors: train b/c nothing better: chosenTrainCommand` |
+
+This reinforces that Train is usually a fallback or quest/prep behavior for the AI, while Human 3 uses it as a primary frontier readiness command.
+
+## Findings To Act On
+
+### 1. Rest Is Still Too Common At Small Sizes
+
+Surviving AI 4-6 at 1-2 provinces:
+
+| Command | Human 3 | AI 4-6 |
+| --- | ---: | ---: |
+| Rest | 1 / 45 = 2.2% | 87 / 466 = 18.7% |
+| Improve | 22 / 45 = 48.9% | 243 / 466 = 52.1% |
+| Train | 0 / 45 = 0.0% | 39 / 466 = 8.4% |
+
+Small AI factions still spend a large share of turns resting. Some of this may be legitimate recovery after pressure, but Human 3 almost never used Rest even while small.
+
+The first follow-up should inspect and adjust `chosenUniversalCommand: rest` and `chosen neutral neighbors: rest`.
+
+Likely rule direction:
+
+- Rest should be a stronger choice only when vigor is actually threatening defense/readiness.
+- Rest should lose to Improve or Train when heroes are merely somewhat tired and enough heroes remain available.
+- Rest should be especially suspicious for small factions with hostile borders or poor strategic position.
+
+### 2. Train Is Under-Prioritized On Hostile Borders
+
+At hostile-border provinces with 3-6 province factions, Human 3 chose Train 26.9% of the time when available. AI 4-6 chose Train 3.9% of the time.
+
+This is probably the highest-value non-Rest difference.
+
+### 3. Divine/Recruitment Pipeline Is Weaker For AI
+
+Recruit itself is fine: AI 4-6 chose RecruitHeroes 100% of the time when available.
+
+The gap is that Human 3 used Divine more often and ended with many more heroes.
+
+### 4. Feast vs HeroGift Still Differs
+
+At hostile-border provinces:
+
+| Command | Human 3 | AI 4-6 |
+| --- | ---: | ---: |
+| Feast | 10.3% | 3.3% |
+| HeroGift | 1.7% | 5.8% |
+
+On hostile borders with 3-6 province factions, AI 4-6 chose HeroGift 10.7% when available, while Human 3 chose it 2.6%.
+
+The earlier Feast/Gift work helped, but AI still appears more Gift-heavy in some comparable contexts.
+
+## Next Step: Rest
+
+Start with Rest because it is both high-volume and likely interfering with better commands.
+
+Questions for the Rest change:
+
+1. Which path produces most of the bad Rest choices: `chosenUniversalCommand: rest` or `chosen neutral neighbors: rest`?
+2. Are these provinces actually under-defended or just moderately fatigued?
+3. When Rest is chosen, were Improve and Train also available?
+4. Should small factions use a stricter Rest threshold unless vigor is defense-critical?
+5. Should Rest defer to Train on hostile borders when enough heroes can still defend?
+
+Desired outcome:
+
+- Keep Rest for genuinely exhausted/defense-critical provinces.
+- Reduce Rest when it is just a low-value recovery command.
+- Prefer Improve or Train when those commands can improve long-term survival without creating immediate vulnerability.