No awards were ever given to the 1989 classic board game Indust and Glonty: Referidon. Holding a solid BGG.com rating of 7, the game can facilitate play from 2 up to 4 players.
No awards were ever given because the game isn’t real.
It never existed in the real world, only in the deep neuron-and-vector tangle of a recurrent neural network. It’s joined by an infinite number games that my computer can now spit out at will, having learned the qualities of over 80,000 board games.
Recurrent neural networks have become the hottest new development in computing, being responsible for the trippy dream-like DeepMind images from Google, generating 3 hours of completely original jazz music and learning how to win Mario Kart on SNES.
I’ll be honest and say I don’t truly understand what’s going inside a recurrent neural network. To be honest many people don’t, which is why they’re so often called ‘magic’.
Two years ago a very smart man called Andrej Karpathy called these network ‘magic’ and provided an example of training neural networks on text. He taught a machine to generate passable-looking Shakespeare and Linux source code that could fool a non-coder.
It takes a huge chunk of text and then attempts to figure out what the next character should probably be. It can then infinitely generate text that looks a lot like huge chunk you gave it – but completely original.
Of course, the ground-breaking technology was crying out to be used on the ground-breaking medium of board games. We’ve combed through the BBG.com database many times before, so we’ve got a bank of over 80,000 board game titles, ratings, details and release dates to feed into the neural network.
After six hours of training on this 4mb text file (!), here’s what the brain-simulating model was able to generate:
Park Glorie (2000) 2-4 players Rating:6 Onth & Gean (1981) 2-2 players Rating:7 Minos's Brin-Mini (2006) 2-4 players Rating:6 Munchkin Park Kings (2008) 2-4 players Rating:6 Flip' El Gays (1964) 1-7 players Rating:4 Power Grid: Fordia (2010) 2-4 players Rating:8 The Besterin Landing: Sentinels of the Alest Tente in the Dark 2 (2001) 4-10 players Rating:5 Secrets! Hall (1988) 2-4 players Rating:6
Blood Bout! (2005) 2-5 players Rating:6 Quesco (2007) 3-5 players Rating:7 Cassin Grachito (2010) 1-4 players Rating:6 Munchkin Tingpor (1992) 2-4 players Rating:6 Nobbit Adventure Card Game (200,) 1-6 players Rating:6 Puzzle Spoin (2013) 2-6 players Rating:6 Krosma (2008) 2-4 players Rating:6 Star Munchkin: Hidernament Hothpek of Bazaday (2007) 2-6 players Rating:6
Flip’ El Gays is the sort of game you’d only be able to get away with in 1964. I could have asked the neural network to generate descriptions, too – but now I’m glad I didn’t.
Some of these game titles are so plausible that I had to check they didn’t exist already. Alas, Quesco, Blood Bout! and Puzzle Spoin do not exist in our universe. In many ways, the neural-network universe reflects our eerily closely – the vast majority of games aren’t highly rated, and there’s far too many versions of Munchkin.
On a few occasions, the network did generate a game that already exists – even if the details are slightly wrong. That’s called overfitting, and makes the output sort of boring:
Buccaneers (2015) 2-5 players Rating:6 Trivial Pursuit (1989) 2-2 players Rating:6 Mamma Mia! (1969) 2-10 players Rating:6
We can make the output even more boring if we want. When the randomness is turned down all the way, the neural network chooses only the most probable set of characters to insert in the title.
Star Wars Miniatures (2009) 2-4 players Rating:6 The Game (2009) 2-4 players Rating:6 Carcassonne: The Card Game (2009) 2-4 players Rating:6 The Card Game (2009) 2-4 players Rating:6 The Game (2009) 2-4 players Rating:6 The Game of Heroes: The Card Game The Card Game (2009) 2-4 players Rating:6 Carcassonne: The Card Game (2009) 2-4 players Rating:6 Star Wars Miniatures (2009) 2-4 players Rating:6
…and the list goes on and on in this manner. I like to imagine a world where there are only three games to choose from: The Game, The Card Card and Star Wars Miniatures. All are mechanically identical and decidedly mediocre.
Turning back the clock is far more interesting. Instead of using the trained model after 6 hours, I trained it again, from scratch, for just 10 minutes on a much smaller set of names.
This is essentially a baby neural network brain. It hasn’t had many experiences, it doesn’t know the cruel realities of the world and it can only come up with board games like this:
Entheyt patty (2001) 2-2 llayers Rating:5 CNel Coocs Zoo U(02) 2-4 players Rating:5 Tcyoonertfaug 3(188772) 2layers Rating:5 TRagy Starrs Matnd inmUponpmal U1Bn Alaccketer Mprs ATi: TdABpaler UMoni atima p:e Bnmale pannid (1974 2- 2laeers Rating:5 Dacey Sttries 2 llanes Rating:4 Hcheon Mnseu (ih2 plalers Ratirn:: ontet Mannes 2iml 2Sllargk Yirt (2000) 2-4 players Rating:5 Poto' Monfess (1071) 2-0 players Rating:5 MDtantaot (2205) --4pplayers Rating:5
After 10 minutes, it seems to have learned that every game needs a rating (and that rating usually isn’t very good). However, it hasn’t quite figured out words yet or how to spell ‘players’. You have to admire the creativity of imagining games made way back in 1071 and also in the near 2205 future, then eighteen millenia from now.
While it’s fun to highlight the silly quirks of the machine-generated board games, I think they can provide the creative seed to some interesting game designs. Some of my favourites were when the algorithm generated titles that were so broad they opened up so many opportunities for board game interpretation:
America (1993) 2-4 players Rating:5 Mexican (2011) 2-5 players Rating:5 World (2005) 2-2 players Rating:6
These sound like a vague seed-word for a 48 hour game jam. What would a board game about America look like in 1993? Would pit the the MTV-era consumerism and counter-culture against each other? Would it a tactical military simulation the Gulf War?
I’m curious to know what imaginary games you come up with given the titles, years and numbers of players in the excerpts above. Let us know in the comments below, or share this article with your imaginary game. A pre-warning though – I’m already in talks with Steve Jackson to create my fully licensed copy of Star Munchkin: Hidernament Hothpek of Bazaday.