the book · contents

04-families 3,817 words about 17 min

Chapter 4. Families

In Factorio the lights are going out. You added a row of assemblers an hour ago, the steam engines cannot keep up, and the grid shares the shortage evenly, so every machine on the map now runs at the same reduced fraction of its speed. The belts back up. Smoke from the furnaces has drifted past the forest that used to absorb it, and a nest of biters at the edge of the map is soaking it up. The pollution it absorbs pays for attackers, 4 units for a small biter up to 400 for a behemoth, and a wave is forming.

In Anno 1800 nothing is on fire. A street of artisans' houses sits one step short of advancing to the next tier. They are full, their basic needs are met, and they have the construction materials, but the canned goulash they want comes from a chain that spans two islands and a ship that is late. When the houses do advance, the artisans who move up stop working in the artisan-tier factories, and a different chain starves.

Both games are production economies. Both are about recipes that turn raw materials into goods through chains of buildings, and both reward the player who gets the ratios right. The dataset still puts them in different families. Factorio sits with Satisfactory, shapez 2, Dyson Sphere Program, and Captain of Industry; Anno 1800 sits with Manor Lords, Banished, Caesar III, and Pharaoh.

What the two games share

Factorio and Anno 1800 share six mechanics: production chains, storage, building placement, blueprints, scheduled vehicle routes, and a large late-game construction project that soaks up surplus. The two most common are production chains, in 38 catalog games, and building placement, in 77.

The parts each game has alone are rarer, and they decide what the game feels like. Factorio's include conveyor belts, a power grid, pollution that spreads chunk by chunk, enemies whose evolution rises with the pollution produced, circuit networks, and robots that move items between chests. Anno 1800's include population tiers whose needs gate their advancement, goods consumed per second by every house, trade routes between islands and a second map, and rival companies. Factorio is a game about throughput, where the challenge is moving items fast enough; Anno is a game about demand, where the challenge is what a growing population will want next.

The grouping method is built to weigh rare parts more heavily than common ones, so it sees that difference.

How the families are drawn

Each game is represented by the set of mechanics it uses. Core mechanics count twice as much as supporting ones. Each mechanic is also weighted by its rarity, the logarithm of the catalog size divided by the number of games that use it, so two games that share a rare part count as much closer than two games that share a pause button. The similarity of two games is the weighted overlap of their sets. The build then groups games by repeatedly joining the most similar groups, measuring similarity between groups as the average over their members, and cuts the resulting tree into a chosen number of clusters.

The weights are steep. A shared storyteller, used by 3 catalog games, carries about 5.4 times the weight of a shared pause button, before the doubling for core parts. On this measure, where 1 would mean identical parts in identical roles, Factorio and Satisfactory score 0.45 and Factorio and Anno 1800 score 0.13. RimWorld scores 0.31 with Dwarf Fortress and 0.2 with The Sims, and SimCity scores 0.2 with Cities: Skylines. Even Factorio and Satisfactory, in one family, share less than half of their weighted parts, because each carries rare parts the other lacks, and those count for the most.

The number of clusters has to come from somewhere. A designer looking at the result could pick the number that gives the neatest story, and the method is set up so that nobody can. The build takes 1000 subsamples, each holding about 80 percent of the games, re-clusters each one, and measures how well each re-clustering agrees with the full one, using the adjusted Rand index, a score that is 1 for perfect agreement and near 0 for chance. The stated rule takes the smallest number of clusters whose average agreement reaches 0.75. When the rule was last applied it picked 24, and that number is fixed in the analysis settings so that adding a game does not reshuffle every cluster.

On today's data the same rule would pick 19. A few games added in October 2026 moved the rule's pick by several clusters, because the agreement curve is nearly flat over that range: every number of clusters from the high teens to thirty scores close to the threshold. The validator warns only when the gap between the configured number and the rule's choice exceeds resampling noise, and it does not today. The number of families is therefore a weak fact. Neighbouring choices are about as defensible as the configured one, so the claims worth making are about which groupings survive resampling.

The book uses two thresholds for that. When it says two games belong together, they share a cluster in at least 0.9 of the subsamples. When it says two games stand apart, they share one in at most 0.1. Anything in between is called a weaker link and pinned by comparison. The build fails if a new game breaks one of these claims. By a separate measure, 15 of the 24 clusters keep an average overlap of at least 0.75 with their best match across subsamples.

Factorio and Anno 1800 share a cluster in about three of ten subsamples. That is a weak link, weaker than Factorio's tie to Satisfactory or Anno's tie to Manor Lords. SimCity stands clearly apart from both, with Cities: Skylines and Pocket City in a family of its own built on zoning, land value, and traffic.

A map of the design space

Figure: A map of the design space. One island per cluster (20 of 24 mix subgenres). Inside an island, lines join games to their nearest neighbours by shared mechanics (inverse-prevalence weighted Jaccard, the measure behind the clustering); bridges join islands whose games are each other's nearest neighbours, wider for more links. Islands sit near the islands they most resemble. A dashed outline marks an island whose games regroup when subsets of the catalog are re-clustered (stability below 0.75). Hover or tap a game to see its neighbours; highlight a subgenre to see it spread across islands. Interactive version: https://sim.yohei.me/#fig-design-map

The map lays out every game near the games it most resembles, with clusters drawn as islands and bridges where many nearest-neighbour links cross between them.

Families that cross genres

Most of the families follow genre lines in part and cross them in part. The crossings are the useful part: each pairs games filed under different genres, and a designer can study them side by side.

Schedules and verdicts. Papers, Please, a 2013 game about checking passports at a border post, belongs with Princess Maker 2, a 1993 game about raising a daughter from age 10 to 18. The resemblance is in the structure. Papers, Please gives you a timed shift, pay per correct decision, and a household ledger of rent, food, and heat that turns speed and accuracy into a moral trade. Princess Maker 2 splits each month into three slots, each spent on training, a job, adventure, or rest, under a small annual income, and her attributes at 18 select one of 74 endings. Both are budgets of scarce time spent against authored events, and the dilemma games Suzerain and Reigns sit in the same group. Princess Maker 2 is labelled a life sim, and it stands apart from The Sims.

Physics you build with. Besiege, The Incredible Machine, and Garry's Mod hold together on rigid bodies, modular parts, and circuits, and two job simulators join them: Job Simulator and Surgeon Simulator. In Job Simulator nearly every object within reach is a physics object that can be grabbed, thrown, or combined, and its variety comes from what players improvise with them. Two other job sims, PowerWash Simulator and House Flipper, stand apart in a family of their own built on surface tools and contracts.

Machines under your hands. Farming Simulator 25 belongs with the driving games, iRacing, Gran Turismo 7, and Euro Truck Simulator 2, and stands apart from the farming games Stardew Valley and Harvest Moon. Its record says why: it models agriculture as ordered field operations with specific implements, across more than 400 licensed vehicles, so the challenge is operating machines and managing a fleet. Stardew Valley's farm is a calendar of crops and a budget of energy.

Rulers and councils. Europa Universalis IV, Civilization, Stellaris, and Tropico 6 hold together on stability, policy choices, and interest groups, across grand strategy, 4X, and city building. Tropico 6 is labelled a city builder and stands apart from SimCity.

Creatures you raise. Little Computer People, Tamagotchi, Creatures, and Black & White hold together on care and learning. Black & White is a god game, and it stands apart from Populous, the game that founded its genre; the creature you train in it puts it in the same family as the virtual pets.

Worlds made of language. AI Dungeon, Infinite Craft, and Where Winds Meet hold together on generated content and natural-language input, and the 2005 interactive drama Façade, which matched typed sentences to a small set of discourse acts, belongs with them. inZOI stands apart from them. It lets typed prompts steer its characters through an on-device model, but its needs-driven core matches The Sims 4, and the clustering puts it there.

Crowds you manage. Theme Hospital, RollerCoaster Tycoon, and Game Dev Tycoon hold together on hiring staff, research, and crowds with preferences. Their visitors and patients have needs as The Sims' characters do, and they still stand apart from the life sims, because what the player manages is a business serving a crowd.

Minds. The life sims hold together, The Sims with Paralives and inZOI, and so do the colony sims, RimWorld with Dwarf Fortress and Oxygen Not Included. The two groups share a cluster in most subsamples, less consistently than either holds together, which fits what chapter 1 found in their parts lists.

Procedure. The flight sims hold together, and three games filed elsewhere sit with them in every subsample or nearly every one: Silent Hunter 3, a Second World War submarine game the catalog files as a vehicle sim, and the wargames Command: Modern Operations and Arma 3. Silent Hunter 3's torpedo solution is in chapter 3; at its hardest realism setting the player also identifies a target's class and nationality, and the crew keeps its rank, experience, and morale from one patrol to the next. Command: Modern Operations simulates the chain from sensor to weapon, so detection depends on radar propagation, terrain, thermal layers, and weather, while the player sets doctrine and missions and lets tactical AI fly and fight. The submarine game shares a cluster with Gran Turismo 7 in under a third of subsamples, and Command stands apart from Hearts of Iron IV. Weather runs through all of them, and the submarine game and the two wargames share military operations and ballistics with DCS World; chapter 9 takes up procedure in the cockpit.

Worlds that keep running. EVE Online holds together with Ultima Online, Second Life, and A Tale in the Desert on persistent worlds, player markets, player organizations, and law that players make. Ultima Online grew skills through practice instead of classes. Second Life has no set objective: its residents build and script almost everything in it and keep the copyright to what they make. Eco has a weaker link, sharing a cluster with EVE in about three of five subsamples. One server holds one set of species, so each player's harvesting and pollution change their numbers and the climate for everyone, and the answer to overuse is the players' own laws, taxes, and currencies, argued with the simulation's data as evidence. Screeps, filed as a programming game, belongs with Second Life, the other world whose residents write code that runs inside it.

Planets in the hand. Populous, SimEarth, From Dust, and Terra Nil hold together on terraforming, god powers, habitat, and water, and the evolution toys SimLife and Species: Artificial Life, Real Evolution join them. SimEarth, built with James Lovelock as adviser, prices its interventions (chapter 2) at 50 units of energy per square of terrain and 500 per terraforming device, and From Dust was designed by Eric Chahi. Terra Nil is filed as a city builder and stands apart from SimCity. It measures growth in restored tiles, pays for buildings only with points earned per restored tile, and ends a map when the player has recycled every building and left. Species, the work of one developer, has no fitness function: genes are numbers, most mutations add a small random amount, and a creature's energy cost comes from the forces and torques in its skeleton.

Programs. SpaceChem, from chapter 2, holds together with Opus Magnum, The Farmer Was Replaced, Bitburner, and Autonauts. In each, the player writes or records a program and the game runs it. Opus Magnum sets no limit on parts and ranks every working machine on cost, cycles, and area against histograms of other players' results, with no reward inside the game for doing better. Bitburner's economy is three calls, hack, grow, and weaken, that move a server's money and security against each other, and the skill is a scheduler that balances them across many servers. Autonauts counts production chains and the consumption of goods among its core parts, and it stands apart from Factorio. Its bots learn by demonstration (chapter 2), and that puts it with the programming games. The Farmer Was Replaced grows crops on a grid and stands apart from Stardew Valley; its drone does nothing unless the player's code tells it to.

Rosters. Football Manager, Out of the Park Baseball, and Motorsport Manager hold together across football, baseball, and motor racing. In each, the player hires staff and sets the instructions that an engine resolves from attributes. Football Manager's database of real players is kept by fan researchers, and Everton agreed in 2008 to use it for scouting. Chapter 9 describes the engine and the board the three share; chapter 5, how two of them hide what a player is worth.

Brushes. Townscaper and Tiny Glade, both filed as city builders, form a family of two that stands apart from SimCity. Neither sets goals or keeps resources. In Tiny Glade a path drawn through a building produces a door, and raising a building adds the columns and beams that hold it up; Townscaper's solver appears in chapter 7.

Bodies in the open. The Legend of Zelda: Breath of the Wild, filed as a sandbox world with survival as a secondary subgenre, belongs with the survival games Subnautica and Don't Starve and with Minecraft. Its record is about its systemic layer rather than its authored map and shrines: a chemistry engine gives objects material properties, so fire burns grass and makes updrafts and metal draws lightning, cold and heat drain health unless clothing or cooked food counter them, and weapons that break and a stamina meter keep the player scavenging.

Series and descendants

Series stay together. SimCity and SimCity 4, The Sims and The Sims 4, Railroad Tycoon and Transport Tycoon, and Anno 1800 and Anno 117 each share a cluster in every subsample or nearly every one. So do Elite (1984) and Elite Dangerous (2014), thirty years apart, on generated star systems, trade, and a ship flown by hand.

The method also finds descendants that share no name with their ancestors. Offworld Trading Company (2015) shares a cluster with M.U.L.E. (1983) in every subsample. Its designer, Soren Johnson, names M.U.L.E. and Railroad Tycoon as inspirations, and the parts show both: a shared market of thirteen resources whose prices move with every purchase and sale, as M.U.L.E.'s auctions set prices between players, and victory by buying up shares in rivals, a mechanic the catalog first records in Railroad Tycoon. The old trading games also have a weaker link to the factory games: Factorio shares a cluster with Railroad Tycoon and with M.U.L.E. in about half the subsamples, and with Satisfactory in all of them.

Among the series in the catalog, Frostpunk is the one that splits.

Frostpunk (2018) is a survival city builder around a single coal generator. Temperature falls on a schedule toward a great storm, the generator heats a circle whose range sets the city's layout, and the Book of Laws offers early compromises such as child labour and extended shifts that open harsher options later, tracked by meters of hope and discontent. Its closest neighbour in the catalog is Against the Storm, another city builder run against a clock.

Frostpunk 2 (2024) moves from single buildings to districts and from one leader's decree to a council. A hundred delegates, each tied to a faction, vote on every proposed law; most laws need 51 votes and laws that expand the Steward's power need two thirds. Promises made to win a vote and then broken anger factions, and tension between factions can escalate to civil war. Its parts put it with Tropico 6 and Europa Universalis IV, and it shares a cluster with its predecessor in fewer than half of the subsamples. The developer describes social survival as the sequel's main threat.

Reading a family

A family is a reading list. Take Frostpunk 2 and its two closest relatives. With Europa Universalis IV it shares interest groups, policy choices, stability and unrest, and a tech tree; with Tropico 6 it shares interest groups, policy choices, and stability, plus elections, but not the tech tree. Europa Universalis IV also has parts Frostpunk 2 does not, and two of them bear directly on a council game. Opinion modifiers make each faction's attitude an itemized sum of named, decaying entries, so a player can read why a group feels the way it does. Causal tooltips, the nested breakdowns of chapter 5, explain any derived number by its named modifiers. Neither record can say whether those parts would suit Frostpunk 2's design. They are the parts its nearest neighbours use to make a political system readable, and a designer working on a game in that family would want to know about them.

The reading also runs from a game filed outside its family. Terra Nil shares a cluster with Populous and From Dust in every subsample or nearly every one. Both make the land the thing the player works on: terraforming is core in each, and in From Dust water and material interactions are core as well, so terrain erodes, floods, and hardens without the player touching it. Terra Nil keeps water and terraforming as supporting parts and counts restoration tile by tile.

The same reading works in any family. Princess Maker 2 ends with one of 74 careers selected by eight years of accumulated attributes; Papers, Please ends with one of twenty, reached by scripted branches from bribes and rule-bending.

How much genre explains

Compare the clusters with the subgenre labels the catalog gives each game. Agreement is moderate by one measure and high by another. The adjusted Rand index between clusters and subgenres is 0.37, where 1 would mean the clusters reproduce the labels exactly. Normalized mutual information, which asks how much knowing a game's cluster tells you about its subgenre, is 0.74. When the subgenre labels are shuffled across games 999 times, mutual information averages 0.42, and no shuffle matches the real labels (p = 0.001).

So mechanics recover genre lineage far better than chance, and they still cross it often: 20 of the 24 clusters contain games from more than one subgenre.

Where the method is weak

The families are as good as the decompositions they are built from, and those have limits that the clustering cannot see.

Games with very few mechanics look alike by what they lack. Conway's Game of Life has 3 mechanics, one of them the turn structure it shares with Hamurabi and The Oregon Trail, and it sits in a cluster with them in nearly every subsample. That cluster is stable and is mostly small early programs; it says little about their designs.

Belonging together is not transitive. Ultima Online and Second Life each share a cluster with EVE Online in at least nine of ten subsamples, and with each other in about four of five; Screeps belongs with Second Life and shares a cluster with EVE in about three of four.

Some clusters do not survive resampling at all. The one that holds Frostpunk, Against the Storm, Plague Inc., Timberborn, and FTL keeps a much lower overlap with its best match across subsamples than the stable families do. Frostpunk and Against the Storm stay together; the others join Frostpunk in some subsamples and not in others.

The clustering also sees only which mechanics a game uses. It ignores scale, tuning, and how the parts are wired together. Two games with the same parts and very different numbers, a gentle farm and a punishing one, look identical to it. And every decomposition was written by one model rater against the taxonomy's definitions, so a different rater could list different parts for the same game. A second, independent set of decompositions would show how much the families move when the rater changes.

Exercise: place a game

Take the game you decomposed in chapter 1.

For each of its core mechanics, note how many catalog games use it. Circle the rare ones.
Find catalog games that share two or more of the circled mechanics. Those are its likely family, whatever their genre.
Pick the member of that family that looks least like your game on its store page, and read its record. List one thing it does that your game could borrow.
Now find a game in your game's genre that shares none of the circled mechanics, and say in one sentence what makes them different.

The families that hold together most tightly include the colony sims, where RimWorld and Dwarf Fortress share a cluster in every subsample. Both carry the top score for emergence in the catalog's rubric. RimWorld's record names Dwarf Fortress as an influence, three of its core parts first appear there, and its record says its systems were chosen for how readable their outputs are: thoughts itemized with sources and durations, injuries named by body part.

Dataset dfde7895aa6d.