Skip to main content Skip to main navigation menu Skip to site footer
Keywords
software, Debord, artificial intelligence, Game of War, Kriegspiel

How I Modeled Guy Debord's Brain in Software

Alexander R. Galloway (New York University)

I was never a member of the cult of Debord. I had read Society of the Spectacle in college, and even watched a few of Guy Debord’s films on my own. I mostly associated Debord with a certain kind of downer avant-garde, all bile and bilge, albeit frequently fun and exhilarating. Later in 2011, now working as a teacher, I assigned Society of the Spectacle in a graduate seminar, mostly for old time’s sake. It bombed. Or maybe the lessons of the book had become so commonplace by then that many younger students didn’t see the point. Why call for a revolution of everyday life, when contemporary life is in constant chaotic rotation? Why call for a form of aesthetic hijacking, when much contemporary art—from memes to games to fine art—is sampled, riffed, and repeated endlessly? Had the work of Debord, that notorious French author and filmmaker and founding member of the Situationist International, finally run its course?1 Had Debord been done in because he won out?

At around the same time, I stumbled on a strange episode from the history of gaming, an episode that would draw me deep into Debord’s archive. I learned that Debord did something rather odd for a leftist avant-garde icon: in January 1977 he started his own game company. Called the Strategic and Historical Game Company, Debord’s new initiative launched with a single game on its roster, a two-player tabletop game called the Game of War that Debord himself had designed and played over the course of several years prior (fig. 1). The company first made a limited edition of the game in copper and silver, then later released a commercial version in paper and cardboard. Collaborating with his wife, Alice Becker-Ho, Debord eventually wrote a whole book about the game.2

Figure 1

Playing Guy Debord's Game of War . (Photograph by Diana Martinez)

Yet Debord’s archive and legacy are complicated. In 1991, a few years before his death, Debord ordered all his extant books to be pulped. The game, already vastly overshadowed by Debord’s more famous books and films, slipped into obscurity, at best garnering a paragraph or two in any biography of this often inscrutable figure.

Many artists have been attracted to games prior to the advent of the personal computer. Yoko Ono made art out of games, as did other artists associated with the midcentury international art movement Fluxus. The Bauhaus reimagined art curriculum around games and playfulness, at least in part. Marcel Duchamp was obsessed with chess. The list goes on. But there was something a bit unusual, even perverse, about Debord’s game. How did this one-time mascot of postwar revolutionary fervor end up devoting part of his later career to a Napoleonic war simulator played by two opposing players on a twenty-by-twenty-five-unit grid?

Given that the game was a kind of allegorical index into the networked and data-driven society growing up around him, I decided to take on Debord’s Game of War as a research project and port the game to the computer. I chose a new name, Kriegspiel, and a new aesthetic style, but kept Debord’s rule set (fig. 2).3 In fact, it was originally a team effort. Mushon Zer-Aviv created the art and graphic design for Kriegspiel, and DJ /rupture and Matt Shadetek wrote a musical score. Daniel Perlin crafted the sound effects. Carolyn Kane and Allison Parrish also contributed to the first version of the game. This new name Kriegspiel—based on the German word meaning “war game”—was in fact Debord’s own informal title for the game, which he used in private letters and notes prior to the game’s official release. Originally built using a Java game engine, the first Kriegspiel software prototype was playable online, with no single-player mode (and hence no artificial intelligence [AI]). Although to be completely candid, I used Debord as a bit of a crutch back then, claiming on principle that Debord wouldn’t have wanted people to play against a computer, when really it was because I hadn’t yet learned how to code AI. Over the years the application got old, and the game became unplayable due to obsolescence. It was also a bit bloated and overdesigned from the outset, with both 2-D and 3-D modes, plus a complicated multiplayer server repurposed from an MMO (massively multiplayer online game). So a few years ago I launched an effort to rewrite the source code from scratch, working on it slowly but diligently during the COVID pandemic. A brand new edition of Kriegspiel was released in late 2021, rethought and redesigned for both mobile and desktop platforms.

What did it take to reenact Debord’s historical algorithm in the present day? What did it take to rebuild this game for mobile devices? Luckily the rules had been published, and there was a decent archival paper trail. (The national library in Paris has even preserved a shoebox full of toy soldiers Debord would use to play army.) So it seemed possible, at least in principle, to rebuild Debord’s game in a new century, in a new format. A number of steps were necessary in developing the game software, including designing the game model, implementing the rules, and adding a networking component for multiplayer. I’d like to tell part of that story here, addressing some details from my attempt to redevelop Debord’s game in software, focusing in particular on the game’s AI. Indeed, the prospect of modeling Debord’s brain was particularly tantalizing. Yet, as we will see, the outcomes were not entirely what I expected at the outset.

Figure 2

Galloway et al., Kriegspiel (2022).

I won’t go deeply into the gameplay itself in this article, but here are the basic coordinates.4 Debord’s Game of War is divided into a northern territory and a southern territory, each with a mountain range, three forts, and two arsenals. The game proceeds in turns. A player may move up to five units each turn, followed by a single attack against an enemy unit. At the same time, each player must maintain lines of communication that criss-cross the terrain, keeping all friendly units within transmission range of one’s home bases. These lines of communication are a special part of the game; they propel player strategy and are key to winning. After the game’s release in the late 1970s, Debord was delighted with his creation. “The surprises of this [game] seem to be inexhaustible,” he confessed later in his memoir. “It might be the only thing in all my work—I’m afraid to admit—that one might dare say has some value.”5

Delighted with his creation, Debord was otherwise disgusted with much of French life in the decades after World War II, whether it was France’s cowardly commitment to Gaullist traditionalism or the craven successes of hypercapitalism. And I have no doubt that Debord, were he alive today, would spit on the world of 2022, oversaturated as it is by spectacle, ideology, and alienation. I also have no doubt he would hate to see his precious army game running on an iPhone. Still, I was never a member of the cult of Debord, and I have no qualms about deviating from the predilections of the master. The Kriegspiel reenactment project is, in a way, a kind of friendly reprisal against this particular historical figure, who, we know, took great pleasure in taking revenge on other people. Rolling in his grave, nevertheless Debord’s restless spirit will rise again as a computer algorithm.

* * *

Nota bene: much of what I will describe below is rudimentary and could be applied to many different kinds of games. This is all material that one might encounter in an introductory textbook on game AI. With that in mind, let’s jump right in and focus on the construction of Debord AI. I’ll add a few thoughts at the end about if and why this matters for Debord in particular.

Figure 3

A node graph connected by links in two different weights, represented by solid lines and dashed lines.

Like chess or go, Debord’s game is highly structured. The game is played on a grid system. Each position in the grid may be occupied by one or more units. Thus, the first task in creating Debord AI was to set up what’s called a graph (fig. 3). A graph is a data structure, which is to say it’s a way of arranging different pieces of data together. The graph consists of a series of nodes. These nodes are connected through various links (also called edges). A simple metaphor might be the streets in a city: plazas and intersections are nodes, which are connected through various avenues and streets. Some roads are one-way only, while other roads are bidirectional. Graphs are very useful. They represent spatial relationships between multiple entities. More generally, graphs represent any form of dynamic relationship. (For instance, a decision tree could also be represented using a graph. The various questions in the decision tree are nodes; answers to questions are links to subsequent nodes in the tree.) So the first step for Debord AI was to set up a graph containing nodes, where each node is a position on the game board able to accommodate the various tokens of the game (arsenals, infantry, mountains, etc.). At the same time, it is necessary to have rudimentary methods for basic utilities relating to the graph, tasks like adding new nodes to the graph, removing existing nodes from the graph, and moving graph elements from one node to another. For instance, if a cavalry unit advances one space, it will be necessary to delete that unit from one node in the graph while adding it to another node in the graph.

Graphs become most useful not simply in static form but when they are dynamically parsed and analyzed. Thus, once the graph is set up it becomes advantageous to move through the graph in various ways. Luckily, computer scientists have devised rudimentary algorithms for doing what’s called graph traversal. These techniques allow a player to move from any arbitrary position in the graph to any other arbitrary position in the graph along the most efficient path. (Graph edges can be given various weights, and these weights can affect the calculation for how much it costs to move through the graph. For instance, a game character might move easily over land but more slowly through water.) Overall, such approaches come under the heading of pathfinding. Pathfinding simply refers to the technique of finding viable paths through a graph (fig. 4). There are many different approaches to pathfinding, some more efficient than others, some more suitable to a particular task than others. For instance, one might wish to start at an origin node and then explore the graph outward in concentric circles (the so-called breadth-first graph traversal). Or one might wish to start at an origin node and then explore as far as possible down one candidate path before trying all the other directions (a depth-first traversal). Breadth-first graph traversals are very common in games for determining proximity. Is the player near an enemy? Or, given one game token, what is the nearest other token?

Figure 4

A pathfinder algorithm showing an efficient path from origin to destination that avoids obstacles.

With a graph established, plus a few rudimentary pathfinding algorithms at hand, one may move game tokens through an imaginary landscape in a somewhat plausible fashion. This allows game characters to target a particular location in the graph, then move toward it. That’s all well and good, but it won’t suffice for most games. The decision-making in Debord AI is more complex than that, and it will require an additional technique. How are targets identified to begin with? Why choose one target as opposed to another target? We need another AI component to supplement pathfinding.

There are many different approaches to decision-making. For complex, open-ended games like first-person shooters, it often makes sense to create what are called behavior heuristics. These are essentially simple rules governed by state variables. For instance, an environmental trigger might move a monster or other nonplayer character (NPC) from an idle state into an aggressive state.

But Debord’s game isn’t like this. The Debord AI necessitates a different approach. The Game of War is not open-ended. In fact the game is rigidly structured, highly rule-bound, and even somewhat deterministic. From an AI perspective, Game of War is more like chess or tic-tac-toe, and thus calls for a specific sort of AI suitable to such games. This kind of AI is called min-max optimization and requires two elements at the outset. First, it is necessary to identify every possible move for any given state of the board game. (For an open-ended game like soccer that involves bodies and objects moving continuously through space, it is very hard if not impossible to itemize every possible move; for a highly structured game like chess, it is trivial to log every possible pawn advance, every possible tour of the knight, every possible translation of the rook, and so on.) Second, one must be able to compute a numerical score for any given state of the game board. To continue with the example of chess: one might count the scores of each piece, giving a relatively high score at the beginning of the match; but if a player loses a queen the player’s score would drop precipitously, and likewise for other pieces. The numerical score itself is not so important. What’s crucial is that any given game configuration could be assigned a specific score, at least in principle. With these two technologies in place—logging all possible moves and assigning each move a score—min-max optimization proceeds as follows: the AI creates an identical duplicate of the game board on which to experiment; the AI executes every possible move in sequence; a score is calculated for the new configuration produced by each move; finally, the AI ranks each move based on its calculated score. With the process of optimization now concluded, one may simply select the highest-ranked move, knowing that this will be the most optimal. (Here, optimal is defined exclusively through the configuration score; the design of such a scoring metric will thus have profound consequences for determining what’s optimal.) In common parlance, min-max optimization is also called brute forcing because it relies on no special finesse or intuition, simply an exhaustive calculation of all possible moves for any given state of the game board.

To sum up the pieces of Debord AI assembled so far: a node graph represents the unit positions on a grid; a pathfinding algorithm determines viable routes through that graph; and a min-max operation collects legal moves and scores them so they can be sorted from best to worst. These tools facilitate a rudimentary level of gameplay. The computer can identify legal moves and, among the legal moves, can identify which moves might be better than others. Yet in this state, the Debord AI is still quite limited. Endowed with a simple syntactical knowledge, it still lacks a larger sense of strategic semantics. Why move into one board territory as opposed to another? How to plan a more complex maneuver that transpires over three or four turns? How to bluff the opponent, or perform a faint?

Figure 5

Data points before (left side) and after (right side) the application of a clustering algorithm such as k -means. In this instance, k is set to 3, generating three distinct groups, marked here as red, green, and blue.

To address these larger questions of strategy, let’s add one last component to Debord AI. Interestingly, graph nodes can be analyzed and filtered like any data set. The kinds of algorithms used in, say, Adobe Photoshop, are often the same used in game programming. They’re just data, after all, and it matters less if the data are a grid of pixels or a grid of game tokens. For this reason, it’s possible to adapt all sorts of data-processing algorithms for use in game AI. Some of the most useful tools borrowed from data analytics are called clustering algorithms. These are simple functions that take structured data as input, then try to identify clusters or groups within the scattered landscape of the data. The k-means algorithm (fig. 5) is one popular way to divide up a data set into various clusters, but there are any number of other ways to do it, optimized for different circumstances.6 For Debord AI, it is useful to know where the enemy units are in the landscape and at what concentration. Is the enemy clustered together in one large group? Or has the group split into two discrete factions? Have any stray units wandered too far away? And likewise for the allied faction: are my allied units concentrated or scattered? Do we have a small or large cluster posted near an enemy stronghold? With the help of a clustering algorithm, the Debord AI is endowed with a bit more strategic knowledge. It can begin to discriminate better between certain maneuvers that might otherwise appear equivalent.

* * *

What does a rudimentary Debord AI say about Debord himself and his legacy? Consider some of the strategies and techniques Debord wrote about, many of which were also characteristic of Situationist International more generally. Recall the techniques of psychogeography and drifting (dérive). Recall Debord’s discussions of plagiarism and hijacking (détournement). Or recall the iconic Situationist map of Paris (fig. 6) in which various neighborhoods are cut up, rearranged, and linked together in a fluid and flexible structure.

Figure 6

Analysis of fluid and flexible movement flows in a Situationist map of Paris. (Drawing derived from Guy Debord, “The Naked City” [1957])

One might naturally assume that Debord’s Game of War would display a similar array of aesthetic values and that the play style would be inductive, associative, rhizomatic, and flexible. One might assume that the game would be about psychogeography, drifting, or hijacking. Surprisingly, Game of War is incompatible with most of these techniques. There is no psychological geography at play in Debord’s game, only a highly rational and deterministic one. And Debord’s play style—which we can glean from the documentation of a match he played with Becker-Ho—is in fact strictly structured into regular shapes that have been optimized for best play (fig. 7).7 My word for it is crystalline because crystals grow by iterating a regular structure. Debord was drawn to regular shapes that have been optimized for the particular rules of the game. These crystalline shapes look like lattices or trusses: crosses, wings, X-shapes, and boxes. Before studying his game, I had thought Debord was a more flexible and dynamic thinker. But after studying the game, I realized he was also logical and deterministic. Maybe Debord was more like a computer than I had previously thought?

Games lend themselves well to the kind of numerical structures and logics native to computers. Not all games, of course; I mean the highly structured and rule-bound games that Roger Caillois clustered under the heading of “agôn” games in his book Man, Play, and Games.8 These kinds of games have scores (numbers). They have clearly defined win states (logical conditions). They have rules (algorithms and control structures). It’s no wonder that these kinds of games work so well on computers. Games already are computers, in a sense—at least certain games already are, Debord’s Game of War included.

Figure 7

Analysis of structured combat relationships in Guy Debord’s Game of War . (Drawing derived from Alice Becker-Ho and Guy Debord, A Game of War [London: Atlas Press, 2007], 33)

We’re left with two possible conclusions, both of which make reasonable sense to me. The first is that computers are less rigid than we think they are. As a data structure, graphs are in fact highly flexible. That famous Situationist map of Paris already looks a lot like a graph. And it wouldn’t be too difficult to apply a clustering algorithm to Debord’s psychogeographic itinerary. (I’m not implying that the results would be fruitful, merely that the material already lends itself to this form.)

The second conclusion is that Debord is not the person that we thought he was. It would be wise not to try to psychoanalyze an author through his works. But the evidence from this particular work is overwhelming. Debord’s Game of War is not about drifting or hijacking. The game is highly structured, rational, and deterministic. In the past I’ve called this a “nostalgic algorithm” given Debord’s affection for a kind of early-modern soldiering that went out with the pith helmet.9 With Debord AI, one might characterize it in another way as well: a tactical algorithm, a way of using code to think about strategy and structure in its most unadulterated form. “I’m not a philosopher,” Debord ultimately confessed during a conversation with Giorgio Agamben, “I’m a strategist.”10

With the benefit of hindsight, we might also say that Debord’s affection for rigid structure means he was (unknowingly, coincidentally) working in the style of symbolic AI—the Good Old-fashioned AI (GOFAI) approach to computer intelligence from the 1970s. This is why I had decided to develop the AI using the min-maxing approach, even though it was an older style of AI: it seemed to match Debord’s structured, deductive style. In essence, Debord was brute-forcing his own game. So it stands to reason that Debord AI should brute-force the game as well.

But I quickly discovered a huge problem with this approach. I learned during beta testing that even the newest, most inexperienced players could beat the computer without trying very hard. As a developer, I make this confession with some dismay, but, alas, my Debord AI just wasn’t that good. Could I figure out a way to make Debord AI smarter? Should I simply keep refining the existing code, hoping to improve it by small increments? Or would I have to chuck the min-max technique entirely and adopt a new approach to AI? After all, today’s AI revolution is driven not by brute-forcing operations so much as data-driven machine learning.

In a sense I was reliving the same kind of crisis in discrete rationality that AI went through in the 1980s, during the so-called AI winter. At the time, symbolic AI just wasn’t good enough to handle complexity. Brute-force rationality wasn’t good enough for real-world scenarios. Maybe min-maxing was okay for chess, but it wasn’t good enough for scenarios requiring more nuanced decisions.

How have computer scientists addressed these problems during the intervening years? And could I follow their example, lest I succumb to my own version of the AI winter? The answer to the first question was that computer scientists changed their entire philosophy toward the very foundations of AI. Instead of brute-force calculations using symbols, laws, and logical deduction, a new kind of AI was developed that was data-driven, statistical, and inductive. Could I, too, leave GOFAI behind, and redesign Debord AI to be more empirical, more data-centric?

Luckily for computer scientists, these historical shifts happened right when users were willing (or compelled) to hand over massive quantities of data, in the form of authored webpages, email and texts, digital photographs, links, likes, and clicks. All this rich data could be harvested by Google or Amazon, analyzed using massive amounts of computing power, and reprocessed into artificial intelligence. The new data-driven approach wouldn’t ever have worked without the data necessary to drive it.

Here I’m left with an ironic conclusion. The old symbolic AI best matches Debord’s play style. But it doesn’t work particularly well as AI, in that it doesn’t make for fun gameplay against the computer. To improve the AI, I suspect I will have to follow the evolution of the computer industry and switch from a deductive/logical approach, to a more empirical, bottom-up approach that relies on data and machine learning. In other words, to best model Debord’s brain, I might need to break with the legible behavior of that brain, which is something I’m still getting my head around. The trouble is I haven’t accumulated enough play data yet to fuel any kind of machine-learning algorithm. And it’s unclear how exactly those data sets could be harvested, and indeed whether they should be. Until that day comes, Debord AI will remain dumb and deterministic, a bit like its namesake, despite his reputation otherwise.

In the meantime, I recommend playing Kriegspiel against a friend instead of the computer. It’s more fun that way.

Footnotes

1. ^ Cofounded by Guy Debord, the Situationist International was an avant-garde group that existed from 1957 to 1972.

2. ^ First published as Alice Becker-Ho and Guy Debord, Le Jeu de la Guerre: Relevé des positions successives de toutes les forces au cours d’une partie (Paris: Éditions Gérard Lebovici, 1987), the book was reissued by Gallimard in 2006, and translated into English as Alice Becker-Ho and Guy Debord, A Game of War, trans. Donald Nicholson-Smith (London: Atlas, 2007).

3. ^ Kriegspiel, updated March 22, 2022, http://r-s-g.org/kriegspiel/

4. ^ While this text is primarily about the Kriegspiel software, I devoted a long chapter to Debord’s Game of War—the game’s history and political context, including some intrigue around Debord’s proclivities as a game designer and player—in my book Uncomputable: Play and Politics in the Long Digital Age (New York: Verso, 2021), 175–212.

5. ^ Guy Debord, Panégyrique, vol. 1 (Paris: Gallimard, 1993), 70.

6. ^ The k-means algorithm filters data into a series of groups, each clustered around a center point or mean. The input variable, k, refers to the desired number of clusters. A typical implementation might look like the following: given a set of one hundred data points, organize these points into four clusters (based on some assumption about similarity, for example, the points’ spatial proximity to one another).

7. ^ See Becker-Ho and Debord, A Game of War.

8. ^ Roger Caillois, Man, Play, and Games, trans. Meyer Barash (New York: Schocken Books, 1979).

9. ^ Galloway, Uncomputable, 197–202.

10. ^ Quoted in Giorgio Agamben, “Repetition and Stoppage—Debord in the Field of Cinema,” in In Girum Imus Nocte et Consumimur Igni—The Situationist International (1957–1972), ed. Stefan Zweifel et al. (Zurich: JRP Ringier, 2006), 36.