DaisenseiMOD

Taking the AI of Civ4 BtS (2007) and turning it into a true “Daisensei” — with every modern trick we’ve got

REPORT CARD

Just getting started…!
Industrial Civ4: done. The AIs can actually play each other
~3.5 sec/game. Reinforcement learning starts here
But: not enough role models (training data) → we need your saves

CURRENT LEVELSETTLER

THREE WAYS TO PLAY

1Show Daisensei who’s bossComing soon!

You’ll be able to play Daisensei head-to-head, right here in the browser. You’ll probably win easily at first. But every game where Daisensei gets crushed becomes study material — and Daisensei keeps getting stronger.

2Toss in your saves and train Daisensei

Old games, this week’s game — bundle one game’s worth of saves into a single folder and drop it in. With a save for every turn, we can see exactly how good players moved, one turn at a time, and that goes straight into Daisensei’s lessons. Duels, free-for-alls, games you lost — all welcome. The more often you toss them in, the faster Daisensei grows.

What to send: A folder with one game’s saves (usually just the auto folder as-is)
Where: Documents\My Games\Beyond the Sword\Saves, inside multi\auto or auto
When: auto gets overwritten by your next game, so send it the same day
Memo: “When & which room” helps us tell games apart
Format: A folder with 5+ .CivBeyondSwordSave files (up to 8 MB each). Modded saves like BUG are fine

What happens to your saves

What we receive: The save files, the folder name, the file dates, and the name and memo you type in. We don’t store IP addresses or other connection logs.
What they’re for: Only for building role models for Daisensei (the AI). We want it to learn strong play, so instead of averaging everything together, we pick games by result and record (the winner, top players, and so on).
Player names: Saves contain in-game player names. To pick good games we may look at who played what. Names are stripped from the data used for training, and nothing is ever published with names attached.
Publishing: The saves themselves are never published. If we ever want to show a game on the site, we’ll hide the names and ask the sender first.
Storage: On the site’s storage (Cloudflare) and on the operator’s machine. Handled only by the operator and the AI helping with the work.
Want them deleted?: Just ping us in the community chat. We’ll find them by name, memo and date, and delete them from both places.

Name Memo

Both are optional — you can leave them blank

Drop your save folder here Drag & drop one game’s folder here, or click to pick a folder

3Watch Daisensei vs. Daisensei — and heckle

Watch Daisensei duel Daisensei. Spot a weird move? Call it out right there.

Watch full screen

Click the tile that bugs you (up goes a smoke signal). For big-picture comments, no click needed
Type one line and hit Enter

e.g. “Pasture this cow”, “Why is the worker building a road here?”, “Way too many settlers overall”. Serious critiques and plain grumbling are both very welcome.

Current games: 1v1 duels (Duel map, Quick speed, Noble), barbarians and huts on, no war until T50, decided on score at T80. Practicing duels first.

ABOUT THIS PROJECT

How far we’ve come

The premiseA strong AI needs a mountain of samples
Make the AIs play each other a lot, and keep collecting samples of “in this position, picking this move got us this much closer to winning.” From those, build a score for every move (an evaluation). The more samples, the sharper the scores, and the better the AI gets at picking good moves. For that, you need a place that can run millions of games.
For scale: chess ~44M games / shogi ~24M / Go ~21M (games AlphaZero played against itself, per the 2018 paper)
The forkThere were two roads
Put a recording rig (a mod) into real Civ4 and make the AIs fight in there. Or build an Industrial Civ4 on the outside, with only the rules an AI needs.
The mod-side recording rig exists too — we use it to record real games.
WhyReal Civ4 can’t do the volume
Real Civ4 is made for humans, so it’s loaded with stuff an AI doesn’t need: graphics, animations, menus. It isn’t just slow — it can’t even get into the ring for millions-of-games sample collecting. So we built a separate Civ4, stripped down to the skeleton of the rules (terrain, cities, production, research, combat, barbarians…). That’s Industrial Civ4.
We check every calculation against the source code the developers released for modders.
DoneIndustrial Civ4 is finished
Feed it a game played in real Civ4, and gold, research, production and population all move exactly like the real thing. ~3.5 sec/game.
On a real game, all 50 turns from T0 to T49 matched exactly. We only use the parts we’ve verified, so the AI can’t get “cheap strong” by exploiting bugs in Industrial Civ4.
You are hereThe AIs have started playing each other in it
Keep what wins, get a little stronger, repeat. Honestly, it’s not going well yet. The games in the spectator seat above come from here.
The problem: not enough role models for the AI to learn from. That’s why we need your saves.

FAQ

Why not just make them fight in real Civ4?: Real Civ4 is built for humans — way too much the AI doesn’t need. To run millions of games you need a skeleton-only Industrial Civ4 (see “Why” above).
Is Industrial Civ4 really the same as the real thing?: We fed in a real game and confirmed 50 turns match exactly. Wherever something differed, we checked it against the developers’ released source code and fixed it one by one. Anything not yet verified (diplomacy, for example) the AI isn’t allowed to use.
So is ChatGPT or Claude playing?: Nope. In-game decisions are made by dedicated parts split by role — economy, combat, barbarian watch and so on — with a coordinator on top. Asking a language model every move would mean thousands of calls per game: way too slow and way too expensive.
How many games are you running?: ~3.5 sec/game. 7 games at once, so roughly 7,000 games an hour.
What do you do with the saves?: They become role models: “in this position, a good player played this next.” The AI starts by imitating that, then sharpens it by playing itself.

What we’re aiming for

An AI you can drop into 1–2 empty seats on a short-handed night and still get a real gameStrong enough not to be your neighbor’s free snack
A capable sparring partner who’s always up for practice or experimentsEspecially for training newcomers
Create a Daisensei that surpasses humans — then learn from itIn chess and Go, people studied the AI’s “moves that look like nonsense at first,” new strategies were born, and the whole top level went up. If that happened in Civ4, it’d be the best

Nerd zone from here (what’s under the hood)

How it’s built

Industrial Civ4 is written from scratch in Python. Checking the math against the source code the developers released for modders, it reproduces the real thing down to combat odds, culture borders, when trade routes update, and research during anarchy
Real games run with a recording mod that writes out the state every turn. We replay that in Industrial Civ4 and check the answers
The spectator seat simply replays games from Industrial Civ4

Why Civ4 is rough on an AI

Too much to decide each turn (production, citizens, workers, units, research and tax rates for every city, all at once)
Results come late. Whether your T10 capital spot or settling at pop 3 vs. 4 was right only shows up around T60
The fog hides the opponent, and combat and huts are luck
Good players decide by feel — “in this position, that one” — and turned into rules, those feelings collide. We tried turning 100+ tips from good players into rules: fix one, and an earlier one breaks. Over and over

What other research says

We haven’t followed all of it, but roughly:

Go and shogi (AlphaZero) beat humans with self-play alone — but those are games where you see the whole board and each move is small
StarCraft II (AlphaStar) started by imitating human replays, then pushed on with self-play up to Grandmaster level. That’s exactly why we want saves
For Civ-likes, there’s a research environment called CivRealm built on Freeciv, and apparently both reinforcement learning and ChatGPT-style AIs struggle a lot
In Civ4, Better BTS AI and K-Mod made the built-in AI much smarter, but it still can’t reach good players. Hand-writing more rules seems to hit a ceiling

Current training

We nudge the numbers inside the AI (when to attack, at what force ratio, how many barbarian lookouts… about 25 of them), play 40 games against the original, keep whichever wins, and repeat
So far no variant has beaten the original. Probably because it gets buried in win/loss luck (around ±8%), and because the biggest strategic choice — “grow ⇔ hit” — isn’t a number yet
Next: search those big strategic choices first → then teach a model, from game data, to predict “how many turns until a win” from the board

Glossary

Industrial Civ4: The AI-only Civ4 built for this project. No graphics, no menus — just rule calculations. Researchers call this a “simulator.” (業務用 = commercial-grade: think restaurant kitchen, not home kitchen.)
Sample: One record of “in this position, I picked this move, and here’s how it ended.” One game yields thousands.
Evaluation: A score for how good a position or move is — an estimate of “how much closer to winning does this get me.” More samples, more accurate. Same idea as a shogi engine’s eval.
Reinforcement learning: Instead of being told the right answer, the AI nudges its evaluations from the results (wins and losses) of its own games. Play → see the result → fix the score, millions of times over.
Self-play: Having the AI play itself. No human opponents needed, so samples are unlimited. AlphaZero beat the best humans at chess, shogi and Go with this alone.
Role models (training data): Records of “a good player played this here.” Having the AI imitate these before reinforcement learning gets it strong much faster than starting from zero. Your saves become this.
Mod: An add-on that changes the game. The Civ4 developers actively supported modding and released source code for it.