Blood Bowl: the ultimate challenge for artificial intelligence?
The best chess engines are now unbeatable by humans.
Go, the ancient Chinese board game considered much more complex than chess, is perhaps not quite solved by artificial intelligence (AI), but AlphaGo, an AI technology developed by DeepMind (a subsidiary of Google), has beaten the top Go grand masters.
Some people are scared of AI because of its potential to somehow eradicate humanity. But me, I won’t be scared of AI until it can beat me in a game of Blood Bowl. Well, maybe not when it beats me (I’m not that great), but someone like Andy Davo.
But what is Blood Bowl, who is Andy Davo, and why is this relevant to AI?
Bring on the Blood
Blood Bowl is a game developed by the British wargaming company, Games Workshop. Set in a parody version of their Lord of the Rings-inspired Warhammer universe, it is a turn based American football-like game played by various fantasy races, such as elves, dwarves and goblins, and characterised by extreme violent play where deaths on the pitch are not only common, but encouraged.
Two teams of 11 players, represented by painted miniature models (aka “minis”) each have 16 turns to score touchdowns. Much like American Football, players can move around the pitch, dodge away from each other, block each other and pass or hand-off the ball. These, and other actions like catching or picking up the ball, are achieved by rolling dice, whereby the dice result required for successfully completing a move depends on the skill of your players, but a 1 is an automatic fail and a 6 is an automatic success.
Depending on the action you are carrying out, failure can mean that your turn end prematurely (a so-called “turnover”). One of the key strategies of Blood Bowl is planning when to take risky actions, versus risk-free ones, depending on their importance.
Under certain circumstances, a failed dice throw can be re-rolled, meaning a critical Blood Bowl coach skill is the ability to work out probabilities of dice results on the fly. e.g. knowing that needing to roll 4 or more on a six-sided dice (D6) with the possibility of a re-roll represents a 75% chance of success (even if it feels like you fail 90% of these!). Or at least having a rough idea…
A major difference from Blood Bowl and almost any other board game that I know of, is that the game is intentionally unbalanced. There are a couple of dozen different races that can be played, and it doesn’t take too much imagination to believe that the Halflings (think Hobbits from LotR) might be at a disadvantage when playing against an agile team of Elves.
There are some mechanisms to slightly reduce the imbalance, but at the end of the day these lesser teams will never win a major tournament. Yet hundreds of coaches will still pick them!
Coupled together with various different “skills” that different players have, or can be assigned which affect the probability of certain actions, the game can get quite complicated. At the end of the day, it is a game of strategy and skill, where (unlike Chess or Go) pure luck can have a massive, and often infuriating, impact on the outcome of any game.
The World Cup
Last weekend I was in Alicante for the Blood Bowl World Cup, where about 2500 fellow Blood Bowl enthusiasts gathered to play this crazy, addictive and often infuriating game. Each game lasts 2.25–2.5 hours, and we played 3 games each day over 3 days.
Over the 3 days, my personal record was W3-D2-L4 — not amazing, but not terrible (losing all 3 games on the middle day was especially grim, though!). There will have been players there who lost all their games — but still had an absolute blast!
What about AI?
Whilst Blood Bowl was born a table-top board game, there is also a computer game adaptation, which is a faithful replica of the tabletop version, with the advantage that you can play online against opponents all over the world, and also that the digital game runs a bit faster (usually <1.25h).
The most recent iteration of the computer game, Blood Bowl 3, was released in February this year after a significant delay. It is universally recognised that the release of the game was a disaster. Nevertheless, after various patches, the game is now more or less playable.
The game’s developers have rightly received a lot of criticism for how bad the release of this game has been, but one area where I feel the criticism is rather unfair is with the game’s in-built AI.
Whilst we take online play against other humans for granted, most newcomers to the game will start off playing in single-player mode, where the opponent is the game’s own AI player.
Anyone who knows the basics of the game, will find the decisions made by the AI somewhat surprising. The AI player will often make ridiculous moves that no human would ever do, which render it very easy to beat. It is no preparation for your first time playing anyone with any experience.
10^30 branches per turn
Understandably, Blood Bowl players lament how badly the AI plays (and the latest version does seem worse than its predecessor). However, I think people really underestimate how complicated Blood Bowl is, or rather they over-estimate how powerful AI can be.
To quote a recent article on the topic:
Every Blood Bowl turn contains lots and lots of tiny decisions that a player won’t normally notice, like the selection of which player to move, the choice of square they move into next, or which action to declare and when. Each of these choices is a ‘branch’ in the turn, and every branch makes it harder for an AI to determine how the game will develop.
To give you a benchmark, there are about 30 branches in each turn of Chess, and 300 in a turn of Go. The AI researchers estimate that, at minimum, a Blood Bowl turn branches the game into 10^30 different possible outcomes — that’s one, followed by 30 zeros. Some of their estimates put another 20 zeros on top of that.
There is an annual AI Blood Bowl competition, where teams pitch their AI models against each other, but even that competition simplifies things by allowing only one type of team to play (Humans).
I’m no AI expert, but I presume you would need a pretty powerful computer to allow it to consider and choose against 10^30 potential moves each turn. Quantum computing anyone?
Forgive the robots
That’s why I do feel a bit bad for some of the criticism the game’s creators get — people, quite understandably, underestimate how complicated Blood Bowl is because they themselves manage to play it without too much bother. We underestimate how much more powerful our brains (even those of Ogre teams coaches) are than computers.
That’s why I won’t be scared of AI until it can beat me, or someone like Andy Davo — one of the game’s most popular online streamers — in a fair, or rather unfair, game of Blood Bowl.
Come and have a go, DeepMind
I’d love to see more AI teams take on the Blood Bowl challenge. I wonder if the DeepMind team had a go at it, whether they would be able to build an Andy Davo-beating AI model? Perhaps Blood Bowl is the ultimate challenge for AI!
