Research: strategic AI arena and living world prior art (persistent faction worlds, bot tournaments, believability, restricted play, fun proxies, off-screen combat models), for rulings arenas-explore-the-metagame and colosseum-feeds-matchup-model
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
soul
parent
2f4b617e50
commit
d52b50f911
1 changed files with 697 additions and 0 deletions
|
|
@ -0,0 +1,697 @@
|
|||
# Strategic AI arena and living world: prior art (Eyes pass)
|
||||
|
||||
Date: 2026-10-21. Eyes, read only. Facts with pointers. No conclusions or recommendations.
|
||||
|
||||
**What this serves.** This pass serves the operator's idea of 2026-10-21, before any fork is put to her. Her words:
|
||||
|
||||
> "You know how we were talking about setting up an arena where we have agents write combat routines and simulate
|
||||
> thousands of matches in various scenarios, from which we can establish various faction combat doctrines (say,
|
||||
> restrict each agent's fleet to a specific faction or set of loadouts)? I just thought, why don't we do the same thing
|
||||
> for the RTS side of the game? That would have way more impact than trying to simulate either the economy via Njordr
|
||||
> or the narrative via Ghostlight, because we already control the game state and all its possible mutations, and when
|
||||
> we have established AI routines for both faction play and combat, we have a living world to present the player,
|
||||
> where the jobs that are available all advance some employer's interests in real time."
|
||||
|
||||
Her scope addition, relayed by Self the same day:
|
||||
|
||||
> "If we're able to score on character, that's where this also takes over Ghostlight's job. And yes, of course the
|
||||
> written routines are content. We run the arenas to surface solutions to the metagame we've established. We run that
|
||||
> in a loop to explore how our games rules and the content we create (tech trees, crafting tree, manufacturing
|
||||
> tradeoffs, combat metagame) cause interesting and fun strategies to emerge. We use agents to explore the metagame
|
||||
> space so that we can iterate on the game design through virtual play."
|
||||
|
||||
**Provenance key.** **[L]** read locally in full: a repo file, the Eureka mind, or a PDF extracted with `pdftotext`.
|
||||
**[F]** a fetched page, read by WebFetch's summarizer model. **[S]** a WebSearch summary only: treat it as a pointer
|
||||
to chase. **[mem]** the agent's prior knowledge, not re-verified in this pass. Aetheria anchors are against `master`
|
||||
`2f4b617e` (2026-10-19), which matches `origin/master`.
|
||||
|
||||
**Earlier passes this one builds on.** They are not repeated here.
|
||||
- `F:\Projects\aetheria-colosseum-prior-art.md` (2026-10-04) covers bot platforms (Robocode, Screeps, Battlecode,
|
||||
Halite, CodinGame, the BWAPI ladders, SC2 AI Arena, Lux AI, CodeClash) and automated balance play in sections 1
|
||||
and 2.
|
||||
- `F:\Projects\aetheria-authored-missions-prior-art.md` covers the X4 Mission Director and the Starsector hub
|
||||
missions (lines 37-75).
|
||||
- `F:\Projects\aetheria-story-sifting-prior-art.md` covers story sifting.
|
||||
|
||||
---
|
||||
|
||||
## 1. Persistent worlds driven by AI factions, and jobs drawn from their needs
|
||||
|
||||
### X3 and X4: Foundations (Egosoft)
|
||||
- **Economy.** Egosoft's marketing says hundreds of stations make wares, thousands of NPC ships carry them, and prices
|
||||
follow that simulated trade. It also says that in X4 every part of the NPC economy is manufactured from
|
||||
resources: ships, weapons, upgrades, ammunition and stations. [S] Steam/press copy via
|
||||
https://dsogaming.com/?p=120136 and https://www.gematsu.com/games/x4-foundations
|
||||
- **Territory.** The Egosoft wiki says factions try to grow their ship and station counts and to annex whole sectors.
|
||||
[S] https://wiki.egosoft.com/X4%20Foundations%20Wiki/Manual%20and%20Guides/Objects%20in%20the%20Game%20Universe/Factions/
|
||||
- **How NPC ships get their work.** NPC ships are spawned to fill per-faction quotas in a `jobs` library file. The
|
||||
`god` file seeds stations, and in X3 the "GOD engine" also kept the economy alive by adding and removing
|
||||
stations. [mem]
|
||||
- **Missions.** Story and generic missions are both Mission Director cue trees. They act on the live universe's
|
||||
sectors, ships and factions, and observe and perturb the economy rather than replace it.
|
||||
[L] `F:\Projects\aetheria-authored-missions-prior-art.md:37-51`
|
||||
- **Documented failures (anecdotal).** Players report early shortages at shipyards and wharfs that leave only one or
|
||||
two stations unaffected. Players also report that war is mostly triggered by the player, and one forum poster
|
||||
calls the economy shallow, with "only one product: ships". [S] https://forum.egosoft.com/viewtopic.php?p=4912608 ;
|
||||
https://steamcommunity.com/app/392160/discussions/0/3104642254774431342
|
||||
- No GDC talk or postmortem on X4's economy was found.
|
||||
|
||||
### Elite Dangerous background simulation (BGS)
|
||||
- **Structure.** Each populated system holds one to seven minor factions, and one of them controls the system.
|
||||
Player actions push factions into states covering economy, security, health and influence.
|
||||
[S] https://forums.frontier.co.uk/goto/post?id=4850407
|
||||
- **Layered states.** Since Chapter Four (2018), a faction can hold several states at once, and a war applies only
|
||||
in the system where it is fought. [S] https://massivelyop.com/?p=192372 . Frontier names political, economic,
|
||||
outfitting and Thargoid layers that interact: Boom improves outfitting and changes prices, and a Thargoid state
|
||||
suspends political movement. [S] same forum guide
|
||||
- **Missions as the lever.** Combat actions and most combat missions push a war toward its result. Elections move on
|
||||
non-combat missions, trade and exploration data. [S]
|
||||
https://www.gamingnexus.com/News/43299/Elite-Dangerous-update-to-combat-issues-that-keep-popping-up-in-Background-Simulation
|
||||
- **Missions draw on faction state.** One player reports that a faction with no state offered no missions.
|
||||
Another reports that missions in Powerplay systems moved no BGS influence. Both are user reports.
|
||||
[S] https://steamcommunity.com/app/359320/discussions/0/1694920442961862234 ;
|
||||
https://forums.frontier.co.uk/threads/missions-not-impacting-bgs-especially-during-elections-in-pp-systems.633066
|
||||
- **Cadence.** BGS states tick daily and Powerplay weekly. [S] same results
|
||||
- **Critique.** Some players say the BGS has little effect on play. Others say it matters for prices, missions and
|
||||
player groups. [S] https://yesxorno.substack.com/p/ed-bgs-and-pp2
|
||||
|
||||
### EVE Online: NPC faction warfare
|
||||
- **Warzone control.** Control is counted as systems held plus installed upgrades, per faction, against the enemy.
|
||||
Rising control raises a warzone tier, which originally scaled LP store prices from 4x dearer to 4x cheaper.
|
||||
[S] https://www.eveonline.com/news/view/fw-dev-blog-1-1-1
|
||||
- **A failure CCP corrected.** Players "spiked" tiers to cash out LP in one day. CCP replaced the store discount with
|
||||
an LP gain bonus. [S] https://www.eveonline.com/news/view/fast-forward
|
||||
- **Tuning.** Victory points from complexes were cut by 5x, which raised the average system capture from 6-7 hours
|
||||
to 30-35 hours. In March 2020, FW mission LP payouts were cut by about 30%. [S] same blogs;
|
||||
https://forums.eveonline.com/t/patch-notes-for-march-2020-releases/224859/213
|
||||
- **Who moves the front.** Players do: the NPC militias are mission givers and site defenders, and do not drive the
|
||||
front themselves. [mem]
|
||||
|
||||
### Mount & Blade II: Bannerlord
|
||||
- **Issues.** Quests are "issues" attached to notables, generated by campaign behaviours from settlement state; a
|
||||
village headman needing grain is one example. [mem] (the `IssueBase` type was not confirmed against the docs this
|
||||
pass). Modders can add campaign behaviours, which run while the player is on the campaign map. [S]
|
||||
https://docs.bannerlordmodding.com/ , https://www.gamebanshee.com/kdyqp
|
||||
- **Documented failures.**
|
||||
- Issue quests are repetitive, and the conspiracy quests send the player across the map during campaigns. [S]
|
||||
https://gamebanshee.com/9k6hq
|
||||
- Bandit quests fail because too few bandit parties spawn. [S] Steam discussions in the same search
|
||||
- TaleWorlds planned to improve "the issue creation algorithm" so players meet a broader range of issues. [S]
|
||||
https://www.gamebanshee.com/383c4
|
||||
- AI armies leave newly won fiefs undefended, and AI lords mismanage conquered towns and lose them to rebels. [S]
|
||||
review and Steam threads
|
||||
- The AI declares war on whichever kingdom it deems strongest, often the player's. [S] Steam threads
|
||||
- Fief-less factions recruiting lords produced large armies with no territory, which players found confusing. [S]
|
||||
|
||||
### Dwarf Fortress: world generation
|
||||
- **What is simulated.** World generation runs history for a set number of years: civilizations, races and religions
|
||||
spread, wars break out, and population and death counters climb. Legends mode exposes every civilization and
|
||||
historical figure. [S] https://dwarffortresswiki.org/index.php/DF2014:World_gen ;
|
||||
https://twinfinite.net/guides/best-dwarf-fortress-world-generation-settings/
|
||||
- **Parameters.** History length, number of civilizations, number of beasts, savagery and maximum sites. [S] same
|
||||
- **Adams on the game.** He calls it a "fantasy world generator" in which the player takes part in a small story.
|
||||
[S] https://scalar.usc.edu/works/game-changers-2016-exhibit/dwarf-fortress . His criterion for adding a
|
||||
simulation (Gamescom 2026) is whether playing through it yourself felt cool. [S]
|
||||
https://www.pcgamer.com/gaming-industry/game-development/dwarf-fortress-creator-says-we-have-a-natural-lean-towards-human-made-stories-but-but-were-all-going-to-be-very-sorely-tested-in-the-coming-years-by-ai/
|
||||
- No primary source on army or site simulation was found. The FDG 2019 and UCSC 2019 emergent-narrative talks are
|
||||
listed at https://www.bay12games.com/dwarves/links.html. [S]
|
||||
|
||||
### Space Rangers 2: Dominators
|
||||
- **AI rivals.** Other AI Rangers compete with the player on a ranked leaderboard, ask for help, and propose joint
|
||||
attacks. [S] https://www.gamespot.com/space-rangers-2-rise-of-the-dominators/user-reviews/2200-489673/
|
||||
- **Economy.** Prices move with supply and demand, local Dominator attacks, and events such as famine. Many events
|
||||
happen without the player. [S] https://sf-encyclopedia.com/entry/space_rangers ; gamesfirst review
|
||||
- The search found no developer documentation.
|
||||
|
||||
### Starsector
|
||||
- **Colony crises (Nov 2023).** Crises replaced punitive expeditions for most factions and are tailored per faction.
|
||||
The developer notes that the old raids by Pirates and the Path were unlimited. [S]
|
||||
https://fractalsoftworks.com/2023/11/24/colony-crises
|
||||
- **Crisis mechanics (wiki, 0.98a).**
|
||||
- Points accrue, preparation starts at 500, and the attack comes at 600.
|
||||
- One-time reductions build "blowback" that feeds 5% back each month, so they only postpone a crisis.
|
||||
- Resolutions can be narrative: a treaty with Kanta, or a deal with the Path.
|
||||
- [S] https://starsector.wiki.gg/wiki/Colony_Crises
|
||||
- **Campaign missions.** Missions locate themselves in the generated sector by require, prefer and pick search.
|
||||
[L] `F:\Projects\aetheria-authored-missions-prior-art.md:53-73`
|
||||
- **Fleets.** The economy is abstract supply and demand per market, and fleets are spawned by per-market managers
|
||||
(patrols, traders) rather than carrying actual goods. [mem]
|
||||
|
||||
### Stellaris AI
|
||||
- **Dev diary #252.** Patch 3.3 aimed at AI economic management in the mid and late game. Patch 3.4 turned to fleet
|
||||
management and diplomacy, and its fleet changes were "mostly bug fixes". The diary opens on the AI splitting its
|
||||
fleets into tiny pieces. Once the AI grew stronger, players objected to it passing Galactic Community resolutions
|
||||
against them. [S]
|
||||
https://admin-forum.paradoxplaza.com/forum/developer-diary/stellaris-dev-diary-252-artificial-and-automated-intelligence.1521980/
|
||||
|
||||
### Ultima Online's original ecology
|
||||
- **Design.** Raph and Kristen Koster designed it: grass grows, herbivores eat the grass, carnivores hunt the
|
||||
herbivores, and dragons pursue a Maslow-like hierarchy of needs. [S] https://en.wikipedia.org/wiki/Ultima_Online
|
||||
- **What Koster says happened.** [F] https://www.raphkoster.com/?p=46439
|
||||
- The ecology ran in alpha. During beta an engineer rewrote it and then the AI was disabled; the resource values
|
||||
stayed and still feed crafting.
|
||||
- It was cut for performance: radial searches followed by pathfinding were too expensive.
|
||||
- He separates the cut from the "collapse". The economy was a closed loop with a fixed resource pool. Players
|
||||
hoarded (sheep killed, wool made into shirts for skill grinding, the shirts held), the central bank ran out of
|
||||
wool, and no more sheep could spawn.
|
||||
- Zach Simpson's analysis of UO's economy is the source of the term "faucet-drain economy".
|
||||
- **Starr Long's account** is that players killed everything faster than it could respawn. [S] Wikipedia
|
||||
- **Wolves.** In 2001 Koster said long-lived wolves grew cannier because there was no cap on how smart they could
|
||||
get. [S] https://ultimacodex.com/2014/07/pimps-and-dragons-ultima-online-profiled-in-the-new-yorker-circa-2001/
|
||||
|
||||
### Pre-simulation, snowballing, and level of detail
|
||||
|
||||
The operator added (2026-10-21): "And since we need all of this to run many times faster than realtime anyway, we
|
||||
can easily simulate a living world for the player when we drop them into the game".
|
||||
|
||||
- **Worlds simulated before the player arrives.**
|
||||
- Dwarf Fortress runs history for a chosen number of years before play, and Legends mode exposes the result.
|
||||
[S] see Dwarf Fortress above
|
||||
- Caves of Qud generates the histories of five ancient Sultans at the start of each run. Grinblat (GDC 2018)
|
||||
avoided a full historical simulation: events are generated first and their causes rationalised afterwards with
|
||||
a state machine and a replacement grammar. [S] https://gdcvault.com/play/1024990/Procedurally-Generating-History-in-Caves ;
|
||||
https://pcgworkshop.com/archive/grinblat2017subverting.pdf
|
||||
- Ghostlight's own storylet ruling has the same shape: simulate to a baseline, then branch. [L] section 4.4
|
||||
- **Runaway snowballing, and how games damp it.**
|
||||
- *Bannerlord.* Early-access patches added relation penalties after raids and sieges and stopped clans defecting
|
||||
to kingdoms that had taken their settlements, explicitly "to prevent players from easily snowballing". Players
|
||||
reported the snowballing persisted. [S] https://www.thesixthaxis.com/2020/04/06/mount-blade-2-bannerlord-crash-fix-update-1-0-5/ ;
|
||||
https://steamcommunity.com/app/261550/discussions/0/2264689848033111449
|
||||
- The AI declaring war on whichever kingdom it judges strongest acts as an emergent coalition. [S] section 1
|
||||
Bannerlord
|
||||
- *Stellaris.*
|
||||
- AI empires snowball through federations: one AI megacorp grew its influence by inviting everyone into its
|
||||
federation. [S] https://steamcommunity.com/app/281990/discussions/0/3826416272371455457
|
||||
- Damping comes through Galactic Community resolutions, the Custodian and Emperor mechanic (declaring Empire
|
||||
dissolves federations), and end-game crises that force cooperation. [S] same thread;
|
||||
https://guildorder.com/tr/games/stellaris/wiki/end-game-crises
|
||||
- Paradox also weights AI aggression against relative power, and difficulty scaling changes AI bonuses. [mem]
|
||||
- *Other catch-up mechanics* [mem]:
|
||||
- EU4: coalitions against high aggressive expansion, and overextension.
|
||||
- Civilization: rising costs as a civ grows.
|
||||
- Starsector: crisis "blowback" only postpones a crisis. [S] above
|
||||
- **Level of detail: full simulation near the player, abstract far away.**
|
||||
- X3: objects out of sector reduce to statistical, turn-based events without collisions. [S] section 7
|
||||
- X4: high and low attention, set by distance from the player. Low-attention combat and movement are separately
|
||||
tuned and patched. [S] section 7
|
||||
- Elite's BGS is a daily, system-level state machine of influence and faction states, fed by player actions. It
|
||||
does not simulate ships. [S] section 1, Elite
|
||||
- Ghostlight's multiresolution gestalts "aggregate simulation work, not canonical identity or state": a villager
|
||||
becomes an individual when attended to and returns to a cheaper population turn while keeping a durable delta.
|
||||
[L via voidbot, `Ghostlight/docs/articles/multiresolution-gestalts.md:28-55`]
|
||||
- Aetheria today ticks only the player's zone; the other zones stay packed. [L] section 4.1
|
||||
|
||||
---
|
||||
|
||||
## 2. Bot tournaments and agent-written strategy
|
||||
|
||||
Platforms and ratings: see `F:\Projects\aetheria-colosseum-prior-art.md` sections 1.1-1.13 (Robocode, Screeps,
|
||||
Battlecode, Halite, CodinGame, BWAPI/SSCAIT/AIIDE/CIG, SC2 AI Arena, Lux, CodeClash). This pass adds the following.
|
||||
|
||||
- **microRTS competition.**
|
||||
- It began at IEEE CIG 2017 with three entries (Ontañón et al., AI Magazine 2018). [S]
|
||||
https://ojs.aaai.org/index.php/aimagazine/article/view/2777/2687
|
||||
- It has been held at CoG since 2019: a round robin on 12 maps (8 open, 4 hidden) with a 100 ms limit per step. [S]
|
||||
https://arxiv.org/html/2402.08112v1
|
||||
- Scripted agents won the first five editions. The first deep RL winner, RAISocketAI, won in 2023. [S] same
|
||||
- **Generals.io.** A public bot server and API exist (`bot.generals.io`, socket.io). [mem; the search returned
|
||||
nothing]
|
||||
- **Screeps.** A persistent MMO of player-written JS bots with no ladder; ranking is by GCL and power. Tick length
|
||||
depends on player load. [L] `F:\Projects\aetheria-colosseum-prior-art.md:25-30`
|
||||
- **LLM agents playing or writing strategy.**
|
||||
- **CICERO** (Meta FAIR, *Science* 2022).
|
||||
- It joins a 2.7B dialogue model to the piKL planner. piKL improves predicted policies toward higher expected
|
||||
value while a KL term keeps them close to human-like play.
|
||||
- It ranked in the top 10% over 40 webDiplomacy games, scoring 25.8% against a human average of 12.4%.
|
||||
- Reported weaknesses: it relies on human data and adapts poorly to out-of-distribution metagames such as
|
||||
variant maps (course slides).
|
||||
- [S] https://ai.meta.com/blog/cicero-ai-negotiates-persuades-and-cooperates-with-people/ ;
|
||||
https://cw.fel.cvut.cz/b251/_media/courses/xp36rgm/diplomacy.pdf
|
||||
- **Voyager** (Wang et al. 2023): an automatic curriculum, a growing skill library of executable code, and
|
||||
iterative prompting from environment feedback and self-verification. Code is its action space because programs
|
||||
express temporally extended, compositional behaviour. [S] https://arxiv.org/html/2305.16291v2
|
||||
- **TextStarCraft II and Chain of Summarization** (NeurIPS 2024): most of 10 LLMs beat the level-5 built-in AI,
|
||||
and fine-tuned LLMs played on par with Gold-league humans. [S] https://arxiv.org/abs/2312.11865v2
|
||||
- **LLM-SMAC** (Deng et al. 2024): a planner, a coder that writes decision-tree Python, and a critic, for StarCraft
|
||||
micro. [S] https://www.promptlayer.com/research-papers/can-llms-master-starcraft-a-new-approach-to-game-ai
|
||||
- **"From Code to Play"** (Eberhardinger et al., arXiv 2412.04057; ToG 2025): LLM-driven hill climbing over game
|
||||
programs on 29 tasks with 11 models. LLM performance depended more on the task than on model size. [S]
|
||||
https://arxiv.org/html/2412.04057v2
|
||||
- **Programmatic strategies for RTS** by script synthesis (AAAI 2021). [S] search result only
|
||||
- **CodeClash** (2025): LLM-written bots.
|
||||
- More than 65% of edits were ungrounded.
|
||||
- Models hallucinated reasons for losses.
|
||||
- Models recovered poorly after a loss.
|
||||
- No model beat expert human bots.
|
||||
- [L] `F:\Projects\aetheria-colosseum-prior-art.md:78-82`
|
||||
- **Known failure modes, with sources.**
|
||||
- *Degenerate exploits.* Battlecode dominant strategies were patched mid-contest (2019, 2021, 2023). OpenAI
|
||||
hide-and-seek agents surfed boxes through physics bugs. Krakovna's specification-gaming list.
|
||||
[L] colosseum file :32-37, :110-112
|
||||
- *Superhuman or out-of-character optimal play.* OpenAI Five used instant buyback, five invincible couriers and
|
||||
death-ball aggression. AlphaStar needed APM and camera limits to be comparable to humans.
|
||||
[L] colosseum file :103-104
|
||||
- *Overfitting to the arena.* Opponent-learning bots dominate long round robins (AIIDE 2017). Halite's local
|
||||
results diverged from the ladder. CICERO adapts poorly to variant maps.
|
||||
[L] colosseum file :50-56, :39-44 ; [S] CICERO slides
|
||||
- *Drift when the ruleset changes.* Battlecode buffs created new dominant strategies (2023 anchoring).
|
||||
[L] colosseum file :37 . EVE's LP tier change was a rule change that players had arbitraged.
|
||||
[S] Fast Forward blog
|
||||
|
||||
---
|
||||
|
||||
## 3. Fitness for believability or interest, not winning
|
||||
|
||||
- **Left 4 Dead AI Director** (Booth, Valve 2009). [L] `pdftotext` of
|
||||
https://cdn.fastly.steamstatic.com/apps/valve/2009/ai_systems_of_l4d_mike_booth.pdf
|
||||
- The stated goal is "Generate dramatic game pacing", inspired by Counter-Strike's "spiky" pacing: constant combat
|
||||
is fatiguing and long inactivity is boring.
|
||||
- It estimates each survivor's "emotional intensity". Intensity rises with damage taken, incapacitation and being
|
||||
pulled off a ledge, decays toward zero, and does not decay while the survivor is engaged. The Director tracks the
|
||||
maximum over the four survivors.
|
||||
- It runs four states. **Build Up**: full threat population until intensity crosses the peak threshold.
|
||||
**Sustain Peak**: 3-5 more seconds. **Peak Fade**: minimal population until intensity decays and a natural break
|
||||
comes. **Relax**: 30-45 s, or until the survivors near the next safe room.
|
||||
- A second goal is "Promote Replayability: Procedurally Populated Environments".
|
||||
- **RimWorld storytellers.** Cassandra Classic, Phoebe Chillax and Randy Random are the player's "unseeable God".
|
||||
Cassandra runs a classic rising curve of challenge and tension. The storyteller weighs colony wealth, colonist and
|
||||
animal counts, recent deaths or injuries, and time since the last major event, and difficulty is set separately.
|
||||
[S] https://rimworldwiki.com/wiki/AI_Storytellers . Sylvester's GDC 2017 talk framed the game "not as a game, but
|
||||
as a story generator". [S] https://gdconf.com/news/gdc-youll-see-rimworld-devs-succeeded-everything-wrong
|
||||
- **Façade drama manager** (Mateas and Stern 2003). It sequences beats from a large pool so that tension rises and
|
||||
falls along an Aristotelian arc. Beats carry author-written preconditions and effects on story state, and the
|
||||
manager adds and retracts behaviours and discourse contexts as play runs. [S]
|
||||
https://users.soe.ucsc.edu/~michaelm/tenurereview/publications/mateas-tidse2003.pdf
|
||||
- **Oz Project lineage.** ABL descends from Loyall's Hap (Bates 1992; Bates, Loyall and Reilly 1992; Loyall 1997),
|
||||
a reactive planning language for "believable agents" that express rich personality. [S]
|
||||
https://course.ccs.neu.edu/cs5150f14/readings/mateas_abl.pdf
|
||||
- **2K BotPrize (UT2004), a Turing test for bots.** Judges tagged opponents as human or bot with a special gun. In
|
||||
2012 UT^2 and MirrorBot were both rated 52% human, while human players averaged 40%. In 2010 the best entry reached
|
||||
31.82%. [S] https://news.utexas.edu/2012/09/26/artificially-intelligent-game-bots-pass-the-turing-test-on-turings-centenary/ ;
|
||||
https://ercim-news.ercim.eu/en84/special/conscious-like-bot-wins-the-2k-botprize
|
||||
- **Procedural personas.** Holmgård et al. used MCTS with evolved heuristics to give agents distinct play styles.
|
||||
[L] colosseum file :101
|
||||
- **CICERO's KL regularisation toward human play** is a published way to keep a strong policy "in character" with a
|
||||
reference distribution. [S] above
|
||||
- **Ghostlight's existing machinery for in-character faction decisions.** See section 4.6.
|
||||
|
||||
---
|
||||
|
||||
## 4. Aetheria's own state
|
||||
|
||||
### 4.1 The strategic or galaxy layer today
|
||||
- **Ownership is fixed at generation.** A zone's owner is the faction whose home zone is nearest, if that zone lies
|
||||
within the faction's influence distance. [L] `Assets/Scripts/ServerShared/Galaxy.cs:327-347` (the assignment is
|
||||
at `:346`). Presence is every faction whose home lies within its influence (`:336-339`).
|
||||
- **Nothing changes the owner after that.** The only writes to `GalaxyZone.Owner` are the generation write
|
||||
(`Galaxy.cs:346`) and the load from save (`Galaxy.cs:77`). [L] `git grep` on master
|
||||
- **`GalaxyZone` holds topology plus owner and presence.** It has no strategic state: no stockpiles, strength or
|
||||
front. [L] `Galaxy.cs:563-576`
|
||||
- **Relationships are player standing per faction.** `Galaxy.FactionRelationships` holds one value per faction, the
|
||||
player's standing (`Galaxy.cs:27`, read at `Entity.cs:1301`), on the ladder Hated through Beloved
|
||||
(`Enums.cs:183-189`). Faction-to-faction `Relations` is only a reserved key 16 (`Corporations.cs:133`).
|
||||
`Allegiance` (`Corporations.cs:121-122`) answers whose gear a faction buys. [L]
|
||||
- **Only the player's zone ticks.** Unity steps the one current zone (`Assets/Scripts/Gameplay/ActionGameManager.cs:1357-1361`).
|
||||
Other zones persist as `PackedContents`. No code advances unloaded zones or the galaxy over time. [L]
|
||||
- **Lore: two layers.** `Aetheria/Game Design/Playable Layers.md` (lines 1-36) is marked "historical design
|
||||
synthesis". It describes a cockpit layer and a corporate strategy layer that "may exchange periodic or regional
|
||||
snapshots rather than share one continuously simulated universe". Corporate play would hand the cockpit layer
|
||||
prices, shortages, patrol density, contracts and influence, and cockpit play would hand back purchases, losses,
|
||||
salvage and completed contracts. `Corporate Strategy Layer.md` is marked "Historical, non-product design
|
||||
lineage". [L via voidbot, AetheriaLore]
|
||||
- **Archived RTS work.** The archived AetheriaEve repo once had a commander/RTS client (`Aetheria.Rts.Web`) and a
|
||||
"Starbridge" commander and pilot design (voidbot hits in `AetheriaEve`). Under the standing rule it is archived
|
||||
and is not a consumer. [L via voidbot]
|
||||
|
||||
### 4.2 Faction play and doctrine today
|
||||
- **Doctrine is a typed record on the faction.** `FactionDoctrine` (`Assets/Scripts/ServerShared/Doctrine.cs:24-57`)
|
||||
holds `EngageOn` (Detection, Identified, Provoked, Trespass, Never), grace, hail and comply speed, plus one
|
||||
`RoleDoctrine` (`:59-74`). Keys 5-8 are reserved for faction-play-2 and 9-10 for faction-play-3 (`:25-26`).
|
||||
`Faction.Doctrine` is key 17 (`Corporations.cs:135-137`). Ruling `faction-doctrine-typed`, 2026-10-03. [L]
|
||||
- **One flight per faction per zone decides.** A flight alone decides its members' targets and tasks, and the agent
|
||||
graph decides nothing (`Assets/Scripts/ServerShared/Agents/Flight.cs:7-15`). Flights are created per faction on
|
||||
first join (`Zone.cs:72-75, 311-313`). [L]
|
||||
- **Patrol is the only task ever constructed.** `PatrolOrbitsTask` is built at `Zone.cs:323`. `HaulingTask`,
|
||||
`Mining`, `StationTowing` and `Survey` exist under `Agents/Tasks/` but no code constructs them. [L] grep for `new`
|
||||
- **Trespass.** It is read from zone ownership and security (`Zone.cs:720-725`; `Entity.cs:1265, 1273`). [L]
|
||||
- **Demo scope.** Faction-play-1 to -3 are demo scope; faction-play-4 (the tender loop) and "procedural regional
|
||||
politics and economy" are full game (`docs/aetheria-release-target.md:141-149`;
|
||||
`docs/aetheria-release-map.md:459-462`). [L]
|
||||
- **The candidate design.** `AetheriaLore/Aetheria/Game Design/Faction Play.md` (Imagination, 2026-10-03, "not canon
|
||||
and not adopted") proposes five levers: NPC behaviour, loadouts, resource distribution, visuals and narrative
|
||||
hooks. Its stance table is at lines 396-411. Its smallest proof asks whether "a playtester, without being told,
|
||||
can describe how each faction fights" (lines 501-506). [L via voidbot]
|
||||
|
||||
### 4.3 Faction territory
|
||||
- **The target.** `docs/faction-territory-target.md` (2026-09-19, "target, operator direction only. No map, no
|
||||
schedule") wants every sector and station to carry a supply-chain role and a colonization era (`:35-39`), a
|
||||
readable territory shape per faction (`:40-45`), and real supply links that the economy and piracy can use
|
||||
(`:49-54`). It requires deterministic generation with no runtime model calls (`:55-56, :60-63, :91`). Its open
|
||||
question at `:86-87` is "What happens to a sector when its owner changes hands". [L]
|
||||
- **The doc's anchor is stale.** Its `Galaxy.cs:340` Markov anchor no longer holds: ruling
|
||||
`faction-curated-place-names` (2026-10-17) replaced name files with `Faction.Places` (`Corporations.cs:146-149`).
|
||||
[L]
|
||||
- **Mind follow-up.** `aetheria-release:follow_up:faction-territory-target` is in force with full-game scope.
|
||||
[L] mind query
|
||||
|
||||
### 4.4 Jobs and missions
|
||||
- **No job or mission generator exists on master.** `Galaxy.cs:261-263` leaves the `StoryProcessor` narrative pass
|
||||
commented out. `GalaxyQuest` and `LocationStory` (`Galaxy.cs:578-595`) are data shells. [L]
|
||||
- **Mind follow-up.** `aetheria-release:follow_up:authored-missions` (in force, full game) wants FreeSpace 2 style
|
||||
authored scenes and missions, with mission logic as C# hooks. [L] mind
|
||||
- **Runtime storyteller.** Ruling `storylets-simulate-sift-lift-weave` (2026-10-03): Ghostlight runs offline to make
|
||||
storylets with pre- and postconditions, and Aetheria's runtime storyteller weaves them. Radio chatter is a
|
||||
delivery surface. [L] mind. The draft is `docs/reactive-world-target.md` on branch
|
||||
`eureka/aetheria-release-reactive-world`, headings at lines 1-146. [L] `git show`
|
||||
- **Lore.** The Action RPG Layer's smallest proof includes choosing work and inspecting "its route, payment,
|
||||
claimant, and known risks". [L via voidbot, `Aetheria/Game Design/Action RPG Layer.md:12-20`]
|
||||
|
||||
### 4.5 Headless simulation and the Colosseum
|
||||
- **The playground map.** `docs/headless-playground-cut.md` (2026-09-17, an Imagination cut map) wants a scripted
|
||||
session run headless through one run-lifecycle owner, one input surface and seeded determinism (`:13-54`). Its
|
||||
finding was that movement and firing were already shared between player and AI, while combat resolution and
|
||||
`CombatState` needed Unity (`:56-100`). Its Cut 2 was re-mapped as fire control (commits `b1c241fb`, `a62d52b6`).
|
||||
[L]
|
||||
- **What is headless on master.**
|
||||
- `RunStart` owns new runs, and every run is a `Scenario` (`RunStart.cs` header; `Scenario.cs` header).
|
||||
- Scenario tests step `Zone.Update` headless (e.g. `tests/Aetheria.Shared.Tests/PursuitScenarioTests.cs:35, 86-88`).
|
||||
- There is no `AetherDb play` command (`tools/AetherDb/` file list).
|
||||
- `ItemManager.Random` is still clock-seeded (`ItemManager.cs:20`), while the galaxy uses the run's seed
|
||||
(`GalaxyStage.cs:218-226`).
|
||||
- [L]
|
||||
- **The Colosseum ruling.** `aetheria-release:ruling:colosseum-is-the-npc-behaviour-lab` (2026-10-04, operator):
|
||||
- The bots are the game's own NPC scripts, written in-repo by agents as compiled C# over a growing primitive
|
||||
vocabulary.
|
||||
- The plan is to "implement all options, and let the good combinations shake out in the arena".
|
||||
- What still matters is a typed observation and action seam, deterministic seeded runs and replays, enough rounds,
|
||||
and exploits treated as bugs.
|
||||
- Losing strategies become products or faction behaviours.
|
||||
- [L] mind
|
||||
- **The Colosseum follow-up.** `aetheria-release:follow_up:colosseum-npc-lab` (in force; owner "Aetheria Agents and
|
||||
the headless simulation") names its foundations:
|
||||
- an observation and action seam (agents hold live `Ship`, `Zone` and `ItemManager` references);
|
||||
- seeding `ItemManager.Random` and the MIConvexHull randoms;
|
||||
- a replay recorder;
|
||||
- a headless tournament runner on Yggdrasil with ratings;
|
||||
- a primitive vocabulary;
|
||||
- a check of float determinism between Windows and Linux.
|
||||
- Why it can wait: "the control rebuild and the drone substrate come first". [L] mind.
|
||||
- The release target lists the Colosseum as not in demo scope (`docs/aetheria-release-target.md:148-149`). [L]
|
||||
|
||||
### 4.6 Njörðr and Ghostlight: what each owns
|
||||
- **Njörðr.**
|
||||
- **What it is.** The GameCult economy daemon for Aetheria and Delvehold: every item comes from a supply chain,
|
||||
materials fill roles, and prices are a pure function of committed state.
|
||||
[L] `F:\Projects\Njordr\README.md:1-14`
|
||||
- **Engine.** A pure, deterministic step function, embedded in the Aetheria client over a C ABI.
|
||||
[L] `AGENTS.md:9-19`
|
||||
- **Status.** The spec reads "specification, 2026-09-11. Nothing is built"
|
||||
(`docs/architecture/njordr.md:3`). The repo is one scaffold commit, `a5bf35e` (2026-09-11), and its
|
||||
`state/map.yaml` says "No step function yet". [L]
|
||||
- **Ghostlight link.** Carrier policies are data: online, a Ghostlight carrier subject owns its policy; offline,
|
||||
seeded policies run. [L via voidbot, `docs/architecture/njordr.md:482-507`]
|
||||
- **Ghostlight.**
|
||||
- **What it is.** A Rust world kernel. A language model proposes, the kernel checks the proposal against typed
|
||||
state, and one `WorldKernel` commits. [L] `F:\Projects\Ghostlight` origin/main `aa56736`, `README.md:1-16`
|
||||
- **Its stated direction.** Storylets for Aetheria, a persistent cast, and episodes for a video stream: "None of
|
||||
this exists yet. Nothing under `crates/` loads Aetheria lore" (`README.md:75-88`). Ruling
|
||||
`ghostlight-verse:ruling:verse-cast-crossmedia` (2026-10-09) names the same three products. [L]
|
||||
- **Strategic wave.** Already specified and documented as a pipeline:
|
||||
- resolution demand over geography, ideology, authority, economy and information;
|
||||
- a deterministic partitioner that builds the active cover;
|
||||
- Nemesis, which picks the subjects that "genuinely need a decision window", possibly none;
|
||||
- per-cell Projector, then Persona, then Interpreter, in parallel;
|
||||
- an effect verifier that stops borrowed authority or knowledge;
|
||||
- an outcome resolver and an outcome verifier;
|
||||
- an atomic kernel commit.
|
||||
- [L] `docs/public-architecture/Runtime pipelines.md:44-62`; stage table at
|
||||
`docs/public-architecture/Agents and model stages.md:111-123`
|
||||
- **Persona's role.** It lets "each required actor, institution, population, or named member appraise and choose
|
||||
independently" (`Agents and model stages.md:117`). [L]
|
||||
- **Ledger limit.** Ghostlight's `Custody`/`Transform` is a unitless, one-to-one ledger, so prices, production,
|
||||
orders and contracts are unrepresentable in Ghostlight. [L via voidbot,
|
||||
`Ghostlight/docs/architecture/delvehold-forced-ontology-integration.md:46`]
|
||||
- **Newspapers.** Agent journalists write over a frozen ledger
|
||||
(`Agents and model stages.md:125-132`). [L]
|
||||
- **What each owns that the arena idea touches.** These are facts, not a recommendation.
|
||||
- Ghostlight owns the following: in-character appraisal by named actors and institutions through Persona;
|
||||
decision windows; durable identity, memory and commitments; prose products (storylets, newspapers, episodes).
|
||||
- Aetheria's game state owns the following: zones, entities, doctrine and flights. The game code applies every
|
||||
mutation (the operator's "we already control the game state and all its possible mutations").
|
||||
- Njörðr owns the following: prices, production, lots and provenance, as a pure step function.
|
||||
|
||||
---
|
||||
|
||||
## 5. Automated playtesting and simulation-driven game design
|
||||
|
||||
- **Jaffe et al., "Evaluating Competitive Game Balance with Restricted Play"** (AIIDE 2012, U. Washington). [L]
|
||||
`pdftotext` of https://cdn.aaai.org/ojs/12513/12513-52-16035-1-2-20201228.pdf
|
||||
- **Measure.** For a restriction R (never play an action, search only k steps, avoid certain outcome states, always
|
||||
pick starting condition s), balance is measured as the probability that the restricted agent A_R beats an
|
||||
unrestricted A that knows R and exploits it (§2).
|
||||
- **Questions it answers.** "Does every action serve a useful role?"
|
||||
- Restricting a piece q no longer lowering win rate flags q as useless.
|
||||
- Greedy versus random versus deep search shows how much the game rewards long-term strategy.
|
||||
- **Why restricted play.** It compares AI and human *skill levels*, not behaviour, because "AI play is not currently
|
||||
predictive of human play" (citing Hingston 2009).
|
||||
- **Case study.** About 20 iterations on the card game *Monsters Divided*, each judged by measures such as
|
||||
"players who never play Red 1/1 win 11.54%".
|
||||
- **Limitations (§5).** Properties "deeply rooted in human psychology, or the look and feel" are out of reach. The
|
||||
prototype used optimal agents on perfect-information discrete games. Hand-coded real-time agents would need
|
||||
adjusting every balancing iteration.
|
||||
- **Tomašev, Paquet, Hassabis and Kramnik, "Assessing Game Balance with AlphaZero"** (arXiv 2009.04374, 2020).
|
||||
- They compared nine atomic rule variants of chess, including No-castling and Torpedo, by training AlphaZero from
|
||||
scratch on each.
|
||||
- Findings: piece values differ between variants, and some variants are more decisive than classical chess.
|
||||
- The paper calls AlphaZero "an alternative in silico means of game balance assessment".
|
||||
- Motivation: heavy opening theory and the high draw rate.
|
||||
- [F] https://arxiv.org/abs/2009.04374 ; [S] ChessBase posts
|
||||
- **Quality-diversity search.**
|
||||
- MAP-Elites with Sliding Boundaries (Fontaine, de Mesentier Silva, Lee, Togelius, Soros, Hoover; GECCO 2019) finds
|
||||
"diverse ways of playing the game well" along chosen behaviour dimensions. It slides cell boundaries to the
|
||||
distribution of evolved individuals, and the paper applies it to Hearthstone design and rebalancing. Its code is
|
||||
`tehqin/EvoSabber`. [S] https://arxiv.org/abs/1904.10656
|
||||
- "Evolving the Hearthstone Meta" searched card changes that equalise deck win rates. Metagame autobalancing tunes
|
||||
rules toward a designer-specified target graph of matchups. [L] colosseum file :102-103
|
||||
- **Browne: Ludi and Ludii.**
|
||||
- *Evolutionary Game Design* (IEEE TCIAIG 2010) defines 57 aesthetic criteria in three groups (intrinsic,
|
||||
viability, quality), measured by self-play, and uses them to evolve games. Its traits include drama, depth,
|
||||
uncertainty and tension. [S]
|
||||
https://gpbib.cs.ucl.ac.uk/gp-html/Browne_2012_sigevolution.html ;
|
||||
https://www.maastrichtuniversity.nl/news/ancient-games-and-artificial-intelligence
|
||||
- Yavalath, one of Ludi's two picks, is ranked in the top 2.5% of abstract board games. [S] same
|
||||
- Four of Browne's measures are restricted-player win rates. [L] Jaffe §4
|
||||
- Ludii implements some of the metrics in its Evaluation module, "not complete and ... not tested extensively"
|
||||
(Browne, 2022 forum). Ludii's current goal is reconstructing ancient games. [S]
|
||||
https://ludii.games/forums/printthread.php?tid=1048 ; https://arxiv.org/pdf/2310.20008
|
||||
- **Machinations (Dormans).** Diagrams of how resources flow through a game's internal economy. The AIIDE 2011 paper
|
||||
"Simulating Mechanics to Study Emergence in Games" added artificial players and charts over many simulated
|
||||
sessions. Micro-Machinations (van Rozen and Dormans 2014) runs the diagrams inside the game software. In
|
||||
Machinations, users script strategies and observe them. [S] https://ojs.aaai.org/index.php/AIIDE/article/view/12477 ;
|
||||
[L] Jaffe §4
|
||||
- **Smith, Nelson and Mateas.**
|
||||
- BIPED, "Computational Support for Play Testing Game Sketches" (AIIDE 2009), lets a game sketch be played by
|
||||
humans or analysed automatically on the Ludocore logic engine. It generates play traces that obey designer
|
||||
constraints. [S] https://ojs.aaai.org/index.php/AIIDE/article/view/12368 ; [L] Jaffe §4
|
||||
- Nelson, "Game Metrics Without Players" (AIIDE workshop 2011), sets out seven strategies for extracting metrics
|
||||
from the game artifact itself. [S] https://ojs.aaai.org/index.php/AIIDE/article/view/12479
|
||||
- **Measures of metagame health in the literature.**
|
||||
- *Intransitivity.* Czarnecki et al., "Real World Games Look Like Spinning Tops" (NeurIPS 2020): games of skill
|
||||
decompose into a transitive strength axis plus cycles that are widest at middle strength, which is why training
|
||||
needs populations of strategies. [S] https://arxiv.org/abs/2004.09468
|
||||
- Balduzzi et al.'s Nash averaging corrects Elo for redundant agents. [L] colosseum file :86
|
||||
- *Strategy diversity.* QD archives (MAP-Elites) count distinct viable styles. [S] above. The 2025 arXiv paper
|
||||
"Beyond Win Rates" clusters character usage for balance in team games. [S] https://arxiv.org/html/2502.01250v2
|
||||
- *Dead content.* Jaffe's "never use action a" measure flags content whose removal costs nothing. [L]
|
||||
- Borovikov's Lanchester-law and rock-paper-scissors metrics apply to unit balance in a 4X. [L] colosseum file :97
|
||||
- *Difficulty distribution.* Isaksen, Gopstein and Nealen (FDG 2015) explored Flappy Bird's parameter space with
|
||||
Monte Carlo play by a human motor-skill model and survival analysis. [S]
|
||||
https://dblp.org/pid/31/1380 ; https://cs.ubc.ca/event/2015/10/exploring-game-space-using-survival-analysis-talk-aaron-isaksen-1
|
||||
- **Cases where AI playtesting misled or fell short.**
|
||||
- AI play does not predict human play (Hingston 2009, cited in Jaffe §1).
|
||||
- Borovikov (EA): agent play was "far from what was observed from human gameplay" unless utility rewards were
|
||||
hand-specified. [L] colosseum file :97, :113
|
||||
- Exploits humans would not find, or could not execute: OpenAI Five's courier and buyback style, AlphaStar's APM,
|
||||
hide-and-seek box surfing, Battlecode mid-contest patches. [L] colosseum file :103-112
|
||||
- Industry survey: developers are skeptical of automated testing agents and want clear objectives and oracles.
|
||||
[L] colosseum file :106
|
||||
- Fun not measured: Jaffe §5 excludes "human psychology, or the look and feel". Browne's aesthetic metrics were
|
||||
fitted to human rankings and remain incomplete in Ludii. [L] / [S] above
|
||||
- Simulated economies diverging from player behaviour: UO's closed loop, broken by hoarding. [F] Koster
|
||||
- **Studio use (from the earlier pass).**
|
||||
- Riot used self-play to find over- and under-powered decks, producing designer KPIs.
|
||||
- Ubisoft's SmartBots run overnight balance reports.
|
||||
- King predicts level difficulty from human-like CNN play.
|
||||
- [L] colosseum file :93-96
|
||||
|
||||
---
|
||||
|
||||
## 6. Computable proxies for fun, and how a human's taste calibrates them
|
||||
|
||||
The operator added (2026-10-21): "Finding the joy in the game is of course down to taste, but I think agents can do a
|
||||
lot of the heavy lifting in a structured joy-finding approach, so to speak".
|
||||
|
||||
- **Browne's Ludi measures.**
|
||||
- Ludi plays each evolved game against itself and scores criteria chosen because human players found them
|
||||
interesting: drama, depth, uncertainty and tension. [S]
|
||||
https://www.maastrichtuniversity.nl/news/ancient-games-and-artificial-intelligence
|
||||
- There are 57 criteria in three groups (intrinsic, viability, quality). [S] section 5 sources
|
||||
- A "playability filter" culls about 98% of evolved games, and the survivors are judged on criteria "previously
|
||||
correlated with human player scores". [S] https://human-competitive.org/sites/default/files/browne-text.txt
|
||||
- The thesis criteria named in the request (lead changes, drama, killer moves, uncertainty) are among them. [mem]
|
||||
Their exact definitions were not retrieved this pass; the source is Browne's QUT thesis, eprints.qut.edu.au/17025.
|
||||
- Two Ludi games were commercially published. Yavalath is the best known, and is ranked in the top 2.5% of abstract
|
||||
board games. [S] same; https://opinionatedgamers.com/2018/04/23/james-nathan-cameron-brownes-automatic-generation-and-evaluation-of-recombination-games/
|
||||
The second is Ndengrod (later Pentalath), published by Nestorgames. [mem]
|
||||
- Browne also proposed elegance and shibui metrics, largely independent of tractability and quality. [S]
|
||||
https://cris.maastrichtuniversity.nl/en/publications/elegance-in-game-design/
|
||||
- **Yannakakis and Togelius: player experience modelling.**
|
||||
- Experience-driven procedural content generation (EDPCG; IEEE Trans. Affective Computing 2(3):147-161, 2011)
|
||||
drives generation from computational models of player experience. [S]
|
||||
https://pure.itu.dk/en/publications/experience-driven-procedural-content-generation/
|
||||
- Player models map gameplay data to annotated experience. One named model type is neuroevolutionary preference
|
||||
learning, used as the evaluation function of search-based generation. [S] https://www.pcgbook.com/chapter10.pdf
|
||||
- The same authors' *Artificial Intelligence and Games* (2018) treats player modelling as one of three pillars.
|
||||
[mem]
|
||||
- Their reported practice is that experience labels are better collected as rank or pairwise comparisons than as
|
||||
absolute ratings (the "ordinal" line of work, with Martínez). [mem; not confirmed this pass]
|
||||
- **Decision density and meaningful choice.**
|
||||
- No peer-reviewed metric that counts "meaningful choices" in a strategy game was found.
|
||||
- Sid Meier's "a game is a series of interesting decisions" (GDC 2012 talk). [mem]
|
||||
- Proxies in the literature:
|
||||
- Jaffe's depth-k restricted players measure how far ahead play must look, and a greedy versus random gap
|
||||
measures how much choice matters. [L] section 5
|
||||
- Browne's uncertainty and drama criteria. [S]
|
||||
- "Choice poetics" (Mawhorter, MIT thesis) is a qualitative unit of analysis for choices. [S]
|
||||
https://dspace.mit.edu/handle/1721.1/118436
|
||||
- Instance complexity measures the difficulty of a choice, tested on knapsack problems, not games. [S]
|
||||
https://www.biorxiv.org/content/10.1101/405449v2.full
|
||||
- **Model judges over play traces.**
|
||||
- "Evaluating Language Models' Evaluations of Games" (arXiv 2510.10930) compares LLM ratings of game designs,
|
||||
made before play, with human ratings. [S] https://arxiv.org/pdf/2510.10930
|
||||
- RPGBench (arXiv 2502.00595) has a judge score interestingness, consistency of facts and of personality in role
|
||||
play, and action quality, checked against a human study. [S]
|
||||
- CausalGame (arXiv 2607.04293) scores agent sessions with a rubric and reports reliability across judge models
|
||||
with inter-rater agreement (ICC). [S]
|
||||
- GameCraft-Bench (arXiv 2606.17861) reruns a multimodal judge ten times on fixed gameplay evidence to test
|
||||
stability. [S]
|
||||
- "LLMs may not be human-level players, but they can be testers": LLM agents' performance correlates with how
|
||||
difficult humans find a game. [S]
|
||||
https://www.promptlayer.com/research-papers/llms-may-not-be-human-level-players-but-they-can-be-testers-measuring-game-difficulty-with-llm-agents
|
||||
- "Online Agent-as-a-Judge" (arXiv 2606.08200) covers believability in life-simulation settings. [S]
|
||||
- **Steering a critic with a few human judgements.**
|
||||
- Christiano et al., "Deep RL from Human Preferences" (NeurIPS 2017) learn a reward model from human comparisons
|
||||
of pairs of trajectory segments. They give feedback on under 1% of interactions (about 0.1% in the NeurIPS
|
||||
listing), with "about an hour of human time" for new behaviours. [S] https://arxiv.org/abs/1706.03741
|
||||
- Sentient Sketchbook (Liapis, Yannakakis, Togelius) models designer style "via choice-based interactive
|
||||
evolution". When the designer picks some suggestions over others, the tool shifts the weight it gives each
|
||||
quality dimension. [S]
|
||||
https://mail.antoniosliapis.com/papers/designer_modeling_for_sentient_sketchbook.pdf
|
||||
- Galactic Arms Race (Hastings, Guha, Stanley) evolves weapons with cgNEAT from players' past preferences, read
|
||||
from use. In a study of over 1,000 online players, players found novel content that also extended weapons they
|
||||
had preferred. [S] https://course.ccs.neu.edu/cs5150f13/readings/hastings_gar.pdf
|
||||
- Browne's criteria weights were fitted to human rankings of games. [S] above
|
||||
- CICERO's piKL regularises a strong policy toward a human-play reference distribution. [S] section 2
|
||||
|
||||
---
|
||||
|
||||
## 7. Statistical combat models that stand in for simulated fights
|
||||
|
||||
The operator added (2026-10-21): "We'll also want to gather combat metagame statistics from the colosseum so we can
|
||||
build a statistical model of matchups to feed the economy simulation with".
|
||||
|
||||
- **Lanchester's laws.** [mem]
|
||||
- The linear law (unaimed or area fire, or one-on-one fights): losses scale with the product of the two forces, so
|
||||
strength is proportional to N × effectiveness.
|
||||
- The square law (aimed fire, where every shooter can engage any target): dA/dt = -βB and dB/dt = -αA, so fighting
|
||||
strength is proportional to αA², and concentrating force pays quadratically.
|
||||
- Empirical tests are contested. A 1975 RAND critique found no clear verification of the square law, and the
|
||||
Dupuy Institute lists tests that could establish neither the square law nor the linear law. [S]
|
||||
https://dupuyinstitute.org/category/methodologies/lanchester-equations/
|
||||
- Borovikov (EA) used Lanchester-law metrics for unit balance in a 4X. [L] colosseum file :97
|
||||
- **Matchup and rating models with covariates.**
|
||||
- Bradley-Terry: P(i beats j) = e^(βi−βj)/(1+e^(βi−βj)). `BradleyTerry2` models ability as a function of
|
||||
covariates, advises adding random effects when it does, and supports contest-level terms such as home advantage.
|
||||
[S] https://rdrr.io/cran/BradleyTerry2/man/BTm.html
|
||||
- Time-varying Bradley-Terry with team factors has been fitted to the Australian Football League. [S]
|
||||
https://arxiv.org/pdf/2405.12588
|
||||
- TrueSkill (Herbrich, Minka, Graepel, NIPS 2006) is a Bayesian factor-graph rating with a mean and variance per
|
||||
player that handles teams and more than two players. [mem] It is used by Halite and CodinGame.
|
||||
[L] colosseum file :39-48
|
||||
- Composition models (Dota 2 drafts):
|
||||
- Draft-only logistic regression reached about 58-70% accuracy.
|
||||
- Conley and Perry's model scored 69.8% but could not capture synergy and counters.
|
||||
- Song et al. went from 58% to 61% by adding hero-combination features.
|
||||
- dota2-predictor uses hero-by-hero synergy and counter matrices, with ROC AUC about 0.65.
|
||||
- Draft-only accuracy is reported as inflated by short matches.
|
||||
- [S] https://arxiv.org/pdf/2106.01782 ; https://cs229.stanford.edu/proj2015/249_report.pdf ;
|
||||
https://arxiv.org/pdf/1701.03162 ; https://github.com/andreiapostoae/dota2-predictor
|
||||
- Per-matchup win rates by race (StarCraft TvZ and the like) are standard community statistics. [mem]
|
||||
- Nash averaging and Elo's blindness to intransitivity are covered in section 5.
|
||||
- **How strategy games resolve off-screen battles.**
|
||||
- **X3 out-of-sector (OOS).**
|
||||
- The community wiki says out-of-sector combat becomes a turn-based system over the full installed weapon power
|
||||
of each ship. It has no collisions, and ships may share a position. [S] https://x3.solasistim.net/index.php/OOS
|
||||
- Turns last 30 s, or 5 s when watched on the sector map; watching did not change the outcome in one forum test.
|
||||
[S] https://forum.egosoft.com/viewtopic.php?p=2474200
|
||||
- Weapons that are carried but not installed are ignored, and only the most damaging missiles matter. [S] same
|
||||
- **X4 "low attention".**
|
||||
- Egosoft now speaks of high and low attention, set by distance from the player rather than by sector. [S]
|
||||
https://steamcommunity.com/app/392160/discussions/0/6664812048260991709
|
||||
- Patch 4.20 "improved combat simulation when player not present".
|
||||
- Later fixes covered capital-ship movement in low attention and stopped main-cannon damage spreading to several
|
||||
targets at once, and station surface-element damage in low attention was reduced by hand. [S]
|
||||
https://steamdb.info/patchnotes/7854799 ; Steam changelogs
|
||||
- **Total War auto-resolve.** Warhammer III adds spells and abilities as a number to a unit's "Combat Potential",
|
||||
because they cannot act in an abstract fight. [S]
|
||||
https://steamcommunity.com/app/1142710/discussions/0/4697910068985751182
|
||||
- **Stellaris.** The Relative Power labels compare fleet power, economy and technology: Superior or Inferior at
|
||||
1.5x, Overwhelming or Pathetic at 2.5x. [S] https://steamcommunity.com/app/281990/discussions/0/4200238856304642295
|
||||
Per-hull weights of 1, 4, 8 and 16 are a player's recollection only. [S]
|
||||
- **Paradox abstractions.** EU4 battles alternate fire and shock phases of three days each, with a 0-9 die per
|
||||
side plus the general's pips, and stacking modifiers. HOI4 compares attack with defence or breakthrough per
|
||||
division. [mem]
|
||||
- **Known complaints when off-screen and on-screen results disagree.**
|
||||
- **X3.**
|
||||
- OOS fights resolve at a 1:1 ratio "so numbers tend to matter more than ship class", and emergency jumps at
|
||||
0% shields were useless OOS. [S] https://steamcommunity.com/app/2820/discussions/0/496880503079542247
|
||||
- Missile and loadout behaviour differs from in-sector fights (forum tests above). [S]
|
||||
- Exploiting the difference (fighting OOS to win fights that are lost in sector) is a common player practice.
|
||||
[mem]
|
||||
- **X4.** Destroyers teleport erratically out of sector, and Kha'ak beams hit harder when the player is away. [S]
|
||||
Steam discussions above
|
||||
- **Total War.**
|
||||
- Manual-battle odds come out far below the campaign estimate (30-40% against 55-60%).
|
||||
- Auto-resolve ignores units that need manual control or specific tactics, and weighs armour and melee too much.
|
||||
- It wipes whole units instead of spreading casualties.
|
||||
- It behaves very differently under stat mods.
|
||||
- [S] https://steamcommunity.com/app/594570/discussions/0/1480982971175660014 and the other threads in this search
|
||||
- **Stellaris.** Fleet power ignores weapon range and composition, so a stronger fleet can lose. An overflow bug
|
||||
once made the total negative. [S] Steam threads above
|
||||
- **How wargame designers fit combat-results tables.**
|
||||
- A combat-results table (CRT) reduces attacker and defender strength to an odds column, and terrain and weather
|
||||
shift the column. [S] https://pushingcardboard.com/base/glossary/1-combat-results-table
|
||||
- Dunnigan's *Complete Wargames Handbook* (1980) has chapters on reading the CRT and on the design process. Its
|
||||
derivation method was not retrieved. [S] https://en.wikipedia.org/wiki/The_Complete_Wargames_Handbook
|
||||
- Dupuy's Quantified Judgment Model fits combat factors to a historical engagement database.
|
||||
- The Joint Staff's 2017 Joint Training Tools reuse its data tables (terrain, weather, water obstacles, advance
|
||||
rates). [S] https://dupuyinstitute.org/category/wargaming/page/21/
|
||||
- The Orion paper rejects Dupuy's "new square law" as a valid attrition model. [S]
|
||||
https://orion.journals.ac.za/pub/article/view/455
|
||||
|
||||
---
|
||||
|
||||
## Gaps
|
||||
|
||||
- No primary developer source was found for X4's `jobs` and `god` mechanics, Bannerlord's issue generator, Dwarf
|
||||
Fortress army and site simulation, Space Rangers 2's AI, or Starsector's fleet spawning. Those lines are [mem] or
|
||||
[S].
|
||||
- No GDC talk on the X4 economy was found. Nothing on the generals.io bot API was retrieved.
|
||||
- The full texts of the Tomašev variant numbers, CICERO's limitations and Browne's 57 criteria were not read.
|
||||
- Frontier documentation for current BGS mission generation was not found; only 2017-2020 community material was.
|
||||
- Browne's exact definitions of lead change and killer moves were not retrieved.
|
||||
- The Lanchester, TrueSkill, EU4 and HOI4 lines are [mem].
|
||||
- No official formula was found for X3 OOS, X4 low attention, Total War auto-resolve or Stellaris fleet power; only
|
||||
community findings were.
|
||||
- How CRTs are derived (Dunnigan, SPI) was not retrieved.
|
||||
- No peer-reviewed metric of "meaningful choice" in strategy games was found.
|
||||
Loading…
Reference in a new issue