zenpundit.com » Blog Archive » Master and (Drone) Commander?

Master and (Drone) Commander?

(by Adam Elkus)

How to think about the shape of future, high-end conventional conflict? Military robotics seems to be a point of recent focus. Take Tom Ricks’ latest on the American military:

By and large, the United States still has an Industrial Age military in an Information Age world. With some exceptions, the focus is more on producing mass strength than achieving precision. Land forces, in particular, need to think less about relying on big bases and more about being able to survive in an era of persistent global surveillance. For example, what will happen when the technological advances of the past decade, such as armed drones controlled from the far side of the planet, are turned against us? A drone is little more than a flying improvised explosive device. What if terrorists find ways to send them to Washington addresses they obtain from the Internet?

Imagine a world where, in a few decades, Google (having acquired Palantir) is the world’s largest defense contractor. Would we want generals who think more like George Patton or Steve Jobs — or who offer a bit of both? How do we get them? These are the sorts of questions the Pentagon should begin addressing. If it does not, we should find leaders — civilian and in uniform — who will.

I quote (as I often do) from John Robb’s excellent analysis of drone swarms because Robb has produced one of the few classics in the emerging military literature on the future of drone warfare. Here, Robb rhapsodizes about the future drone swarm commander and his unlikely origins in the civilian (and South Korea-dominated) Starcraft game series:

Here are some of the characteristics we’ll see in the near future:

  • Swarms.  The cost and size of drones will shrink.  Nearly everyone will have access to drone tech (autopilots already cost less than $30).  Further, the software to enable drones to employ swarm behavior will improve.  So, don’t think in terms of a single drone. Think in terms of a single person controlling hundreds and thousands.
  • Intelligence.  Drones will be smarter than they are today.  The average commercial chip passed the level of insect intelligence a little less than a decade ago (which “magically” resulted in an explosion of drone/bot tech).  Chips will cross rat intelligence in 2018 or so.  Think in terms of each drone being smart enough to follow tactical instructions.
  • Dynamism.  The combination of massive swarms with individual elements being highly intelligent puts combat on an entirely new level.  It requires a warrior that can provide tactical guidance and situational awareness in real time at a level that is beyond current training paradigms.

Training Drone Bonjwas

Based on the above requirements, the best training for drones (in the air and on land) isn’t real world training, it’s tactical games (not first person shooters).  Think in terms of the classic military scifi book, “Ender’s Game” by Orson Scott Card. Of the games currently on the market, the best example of the type of expertise required is Blizzard’s StarCraft, a scifi tactical management game that has amazing multiplayer tactical balance/dynamism.  The game is extremely popular worldwide, but in South Korea, it has reached cult status.  The best players, called Bonjwas, are treated like rock stars, and for good reason:

  • Training of hand/eye/mind.  Speeds of up to 400 keyboard mouse (macro/micro) tactical commands per minute have been attained.  Think about that for a second.  That’s nearly 7 commands a second.
  • Fight multi-player combat simulations  for 10-12 hours a day.  They live the game for a decade and then burn out.   Mind vs. mind competition continuously.
  • To become a bonjwa, you have to defeat millions of opponents to reach the tournament rank, and then dominate the tournament rank for many years.  The ranking system/ladder that farms new talent is global (Korea, China, SEA, North America, and Europe), huge (millions of players), and continuous (24x7x365).

That’s the tactics—but what about the strategy? Robb calls it a “tactical management game,” which is correct. We can discern a bare shell of the “strategy” we normally discuss in the higher level decisions concerning the composition and deployment of the force. And here we also see a different kind of strategic control at play, one much more having to do with the Cold War science of operations research.

One important cognitive aspect of Starcraft that has been automated is the evolution up the tech tree. The tech tree that the player must advance up in order to produce needed units, accessories, and tactics is deterministic, perhaps reflecting the real-world convergence toward a “modern” style of high-end conventional tactics. Starcraft as a game represents the purely tactical considerations of warfare as an elaborate game of rock-paper-scissors, in keeping with Clausewitz’s statement that tactics can be considered closer to science than other aspects of warfare.

It is a reflection of Starcraft‘s deterministic structure that the tech tree “build orders”, the most crucial element of Starcraft‘s mode of war, can be automated. A genetic algorithm infamously was derived to optimize build orders. But this is only possible because the build orders themselves optimize a very small piece of the overall problem, and one made possible by determinism baked into the game.

The use of genetic algorithms to produce build orders also interestingly enough mirrors the overall social, economic, and organizational structure that produces a champion Starcraft player. In the 1980s, Robert Axelrod created an algorithm tournament designed to find a best-performing strategy to the canonical “Prisoner’s Dilemma” in game theory. Using the tournament selection mode of genetic algorithms, Axelrod iteratively weeded out “unfit” strategies until a dominant strategy was found. Perhaps the process that Robb describes is quite literally “tournament selection” that produces an optimal Starcraft player type.

The most important element of strategy — translating organized violence into political payoff — is mostly absent. Starcraft demands the intricate steps needed to prepare the weapon itself (build older optimization) and immaculate skill at firing it (in-game command) but not the problem of ensuring that the violence make political sense. There is no security dilemma caused by the threat of Zergling rushes. 🙂

Because it is a videogame, Starcraft as experienced by the player is nothing close to the overall difficulty, uncertainty, and complexity implied by the overall in-game universe of factions, technologies, and personalities. The level of cognitive difficulty that must be dealt with is kept on the order of something that a single player can reason through. Of course, in even in the “closed” world of real Cold War military science (which Starcraft has eerie similarities to), this has been the stuff of military staffs, RAND and Hudson-like research groups, systems analysts, and supercomputers.

What about uncertainty and complexity? Depending on the game, the most important political-military decisions may not be up to the player. The transformative in-game decision to rebel against Arcturus Mengsk and create Raynor’s Raiders is not made by the player but by the grieving Jim Raynor.  In Starcraft: Brood War and Starcraft II, player choice becomes important in structuring the flow of action. When attacking Char in Starcraft II, the player must choose to either attack the enemy’s air support or ground elements. Both choices are presented are potentially valid depending on player preference. Many other individual choices lead to important distinctions in the shape of events. But the overall “basins of attractions” built into the game structure pull the player towards the same broad outcome regardless. That’s because the game universe and the creators’ demands is the overarching political-military context that determines the path of the war.

When it comes to multiplayer matches, online games in general make combat sport. That is why we dub the Korean Starcraft aces champions. They compete in a ritualized game with clear rules and all-powerful human gamemasters that create the game itself and instantiate their ideas of what an ideal combat sport represents in computer  code. Starcraft has much more in common with the Roman coliseum battles than the Roman army on campaign in some harsh European or Middle Eastern land. Of course, all online environments have weak points that are often exploited to offer advantage, but Starcraft‘s limited range of behavior makes it easier for game-masters to secure than the sprawling World of Warcraft or EVE Online. 

Though I have some serious misgivings about the ethical context of Ender’s Game as a novel, it also remarkably approximates the experience of game-playing in many real-time “strategy” games like StarcraftEnder himself, whom Robb analogizes, is a virtual virtuoso that spends most of his time in Ender’s Game unaware that the “training” simulations he is playing are actually the war he is training to fight in the first place. Hence one comes to wonder if the real genius is not necessarily entirely Ender, who supplies the cognitive firepower necessary to dominate Clausewitz’s “play of chance” on the battlefield. Rather, what about the men and women who organized  and equipped the fleet?  And of the politicians and generals that decided  the overall shape of the strategy that Ender executes, and infamously decided to authorize the genocide of the “Bug” aliens Ender exterminates with weapons of mass destruction?

This isn’t a strike against Robb’s idea that Starcraft is a metaphor for one part of future warfare. Robb himself states that Starcraft is tactical management, and it is as good an vision to contend with as any other. Changes in warfare that begin on the level of tactics have strategic implications. We already know that tactical virtuosity that might be so essential to victory in a closed environment with well-formulated rules are often counterbalanced by the problem of making those skills serve strategic effectiveness outside that environment. What kind of problems might arise for the hypothetical Starcraft-ish military bot commander?

The first problem to be surmounted is collective action. Multi-agent systems face similar coordination problems as seen in human relationships. The interdisciplinary field of algorithmic game theory has arisen to study how to create algorithmic mechanism design for solving many of these issues. Another problem lies in the conflict between speed of tactical execution and the slower-moving demands of strategy. The Cold War stories of commanders that decide to risk annihilation rather than launch nuclear forces on faulty signals tells that many strategic problems have to do not necessarily with the most efficient ways of employing violence but rather have to do with the control of military power. This question has in fact dominated most discussion about autonomous weapons.

Lastly, the most important insight that Robb’s piece gives us is that Starcraft is an social environment that produces novel behavior. It is the online wargaming medium itself and its speed and essentially social complexity that produces the Starcraft champion’s unique characteristics. Similarly, a certain Corsican arose from the cauldron of the “multi-player interaction” of an era caught between the emerging crest of “modern” warfare and the 18th century military system. Dubbed the “God of War,” he became the template for every 19th century commander to copy. The most important strategic problem implied by Robb’s blog is conceptualizing the range of behaviors produced by the unique military system that he sketches with Starcraft as inspiration.

5 Responses to “Master and (Drone) Commander?”

  1. carl Says:

    Drones are exciting aren’t they?  They will be unstoppable and able to do all things at all times.  We should be very afraid.
    Well not so fast.  Airborne drones are little airplanes.  Little airplanes don’t do so well in snow, rain, wind, sandstorms etc especially when they are close to the ground.  They crash a lot.  So they will be able to do some things that can’t be done now but until they overcome the God’s unfavorable winds, maybe they ain’t so fearsome.  And just imagine if you could train crows to knock down little drones.
    Second, drones require radio signals to work.  They are essentially little RC airplanes.  If you can send a radio signal to your drone, somebody else can.  Or somebody else can block it.  Until that problem is overcome, again, maybe they ain’t so fearsome. 
    I like the anti-drone crow idea though.

  2. zen Says:

    Interesting post, Adam.
    I wonder how AI will learn strategy? Much like brute force supercomputers  crunching moves to play chess against human grandmasters within a framework of known rules or via “rulebreaking” experimentation via remote drones and swarms?

  3. AdamElkus Says:

    Mark, machine learning is a pretty complex subject, in large part because of the definition of “learn” that we use for machine intelligence is very different from the colloquial idea of learning. For example, the difference between reproducing a copy of some real-world function vs. correctly predicting. This is a good primer: http://egtheory.wordpress.com/2013/06/05/prediction-vs-understanding/

  4. zen Says:

    Hi Adam,
    This is all very far from my field but a few observation/questions: 
    ” This choice makes the algorithm itself the hypothesis and thus the hypotheses class is equivalent to all efficiently computable functions.
    This seems self-referential. 
    This is how a computer scientist can operationalize prediction and understanding. In a classic result, Pitt & Valiant (1988) showed that there are certain concept classes that cannot be efficiently properly learnt, but that we can learn to predict. In other words, it really is easier to predict than to understand.

    This reminds me of Newtonian mechanistic theoretical breakthroughs in physics – we can predict effects of gravity with mathematics and some data, but we don’t know what gravity actually is. It also reminds me of arguments about the limits of knowing using logic and math

  5. Grurray Says:

    Brute force hasn’t gotten us much more than parlor tricks

Switch to our mobile site