Royale
What a Japanese Cult Classic can teach us about the power of information
If you are short on time, just scroll down and read the last section called “Deception.” It will take you about 30 seconds.
Welcome to The Distro!
The Distro is a military-themed newsletter designed to help make leaders and organizations more effective. Newsletter topics range from decision-making, strategy, risk, uncertainty, command, organizational theory, military operations, history, and leadership. I explore these topics through personal stories, ideas from books, lessons from history, military doctrine, and fictional stories with made-up characters.
Choose Your Weapon
Before there was The Hunger Games there was Battle Royale. I cannot explain to you how incredible this movie is, and Quentin Tarantino has frequently named it as one of his favorite movies of all time. Needless to say, the movie is exceptionally violent.
A quick synopsis: In a dystopian future, a bunch of high school kids are put on an island and forced to battle to the death until there is only one survivor. Each of the contestants is forced to wear a collar. If there is more than one contestant alive at the end of 72 hours, all the collars will explode, and everyone will die.
All the contestants begin in one room. Their names are called one by one. When a contestant’s name is called, they are given a bag with some food and water and a separate bag with a randomly assigned weapon. Some of the weapons that are distributed include guns, bows and arrows, poison, axes, knives, a bullhorn, samurai swords, and a bunch more random stuff. One person is even given a trashcan lid.
The weapons are assigned randomly, but which weapon would you choose? You would choose the machine gun, wouldn’t you?
Knowing is half the battle (Killin’ is the other half)
In the movie, one weapon stands out as clearly the most powerful: a GPS tracker that gives you the location of everyone on the island. Armed with this information, you can develop options and develop a systematic approach to victory. If the goal is to be the last one standing, a good strategy would be to stay alive by staying away from everyone and letting them kill each other for the first 48 hours, and then using the last 24 hours to stalk and kill the survivors. This strategy is only possible if you know where everyone is. Of course, this strategy would not make for a very interesting movie, but it’s probably one of the best strategies in reality.
The only difficult part of this strategy would be acquiring a weapon. But with so many people killing each other, you are bound to come across a few weapons as you move around the island. Because you can track everyone’s movements, it would be relatively easy to set up ambushes and lure people into traps.
One should never underestimate the power of knowing where all the bad guys are. In war, it matters less which side has the best weapons and more which side has the best intelligence and knows how to turn that intelligence into action. If you know where the bad guys are, but they don’t know where you are, the element of surprise is on your side.
Kriegspiel Reprise
Last week I published an essay about a chess variant called Kriegspiel. In Kriegspiel you use all the regular rules of chess, except the players cannot see each other’s pieces — a referee ensures that only legal moves are made. I suggested in that essay that AI may become dominant over the best human players, but that the difference between the strength of the AI and the strength of the best human players would be narrower than for traditional chess where the computer knows where all the pieces are. This is because the uncertainty about the position of the pieces on the board makes it possible for the computer to miscalculate and make mistakes. Trying to program the computer to make moves under uncertainty is not the same as programming the computer to make moves in conditions of certainty.
A lot of people took issue with this in the comments, arguing that AI is so smart and so amazing that it will achieve the same level of dominance over the best human players as it has achieved in traditional chess. But if that were true, why hasn’t it been done yet?
It’s not as though no one has tried.
Several academic researchers have built engines that play Kriegspiel. As far as I can tell, the best attempt was by Dr. Paolo Ciancarini at the University of Bologna whose “Darkboard 2.0” did exceptionally well but was unable to consistently beat the strongest humans. Here is a slide from a presentation Dr. Ciancarini gave in Paris in 20131:
In their paper, Dr. Ciancarini and his team explain that they are very bullish on improvements that can be made to their programming methodology, and express optimism that eventually a Kriegspiel AI could consistently beat the best humans. However, it does not appear that anyone has attempted or succeeded at building a dominant Kriegspiel program. So, if the dominant Kriegspiel AI that I am assured will appear is going to appear, it has some significant catching up to do. AlphaZero and StockFish will continue to improve faster than humans can improve. This means that a Kriegspiel program must have several incredible breakthroughs just to catch up.
Computers starting reliably beating the best humans in 1997 (when Deep Blue beat Kasparov). But a Kriegspiel program has yet to reliably beat the best humans. This means that the Kriegspiel programs are move than 20 years behind if they expect to reach the same level of dominance as traditional chess engines.
To understand why it is so hard to make a strong Kriegspiel program, let’s look at the methodology that Dr. Ciancarini and his team took to develop Darkboard 2.0.
The next right thing
In their paper Monte Carlo tree search in Kriegspiel2, Dr. Ciancarini and his team try three different approaches to developing a Monte Carlo Search Tree (MCST) program to play Kriegspiel. The first two approaches were similar to how traditional chess engines work, they tried to simulate the game all the way to the end and then make the move that most likely led to a victory. But they found that the best approach, called approach “C” in their paper, was to simulate only the next move. In approach C, the simulation stopped after only one move. In this approach they were not trying to find the optimal strategic move for the long term, they were just trying to find the next best tactical move. In other words, the level of uncertainty meant that they had to train Darkboard 2.0 to do the opposite of what traditional chess engines do.
To be fair, approach “B,” which was a longer simulation on just the computer’s own pieces (rather than trying to guess the position of all the pieces), was improving steadily over time, and they suspected that it might eventually surpass approach “C.” But to my knowledge, they didn’t publish additional work on this approach.
The interesting thing is that the best Kriegspiel program, unlike its cousin in traditional chess, essentially mimics the way humans play Kriegspiel. Traditional chess engines take a superhuman human approach and dominate humans, whereas the best Kriegspiel program mimics the way the best humans play and is very competitive, but noy yet dominant.
I say again: I think that, eventually, AI will likely be able to beat the best humans at Kriegspiel consistently. But the difference between the Kriegspiel engines and humans will always be smaller than the difference between traditional chess engines and humans. This is because the element of uncertainty means that the computer will inevitably make mistakes. If it can make mistakes, then it can create opportunities for a human player. Perhaps the human player will make a few good moves by pure chance which gives the human a big enough advantage to win the game.
But let’s elaborate on the power of information, and why information in Kriegspiel (or lack thereof) makes such a big difference.
Blind AI
As I said in the previous essay, AI is only as good as the information that it is given. If it is given bad data, it will produce bad results. If it is given partial information, it will do its best to guess, but it can be wrong.
What if we took our lesson from Battle Royale (that the GPS tracker is the best weapon) and applied it to Kriegspiel?
Imagine a situation in which an AI is playing Kriegspiel but its human opponent is playing regular chess. That is, the human can see the board and all the pieces, but the AI can only see its own pieces and is governed by the rules of Kriegspiel. There is no future in which even the most powerful AI will be able to defeat a strong human player (around 2000 ELO). It does not matter how good the computer is, the fact that the human can see the pieces and the computer can’t is an information asymmetry that is too great to overcome. If you can’t accept this, I cannot take you seriously, sorry.
If I were to play Magnus Carlson where I can see all the pieces but he is constrained by Kriegspiel rules, I am 100% certain that I would beat him.3 There is over a 1,000 point ELO difference between us, but if I can see his pieces and he can’t see mine, he’s at an impossible disadvantage. I’ll use my knights and bishops to capture his queen and rooks, then I’ll use my pawns to constrain his bishops and knights and then I’ll cut open his back ranks with my rooks and queen to get after his king. And there is no way he can stop this from happening, because he can’t see my pieces!
To understand this, you have to appreciate just how hard of a game Kriegspiel is. You might make 5 moves and have gathered no data on your opponent, nor she on you. Imagine you are the white pieces and you make your first move—
Okay, black has moved now, it is now your turn again.
What are you going to do?—
…A sensible move, indeed.
Okay, black has moved, your turn again—
Develop your 2nd knight? well done…Oh No! Sorry, your knight is gone! what are you going to do now?—
What captured it? Probably a pawn, a knight, or a bishop.4 But you have no way of knowing for sure. It should be noted here that the rules I grew up with are different than the rules used by Dr. Ciancarini and his team. The way I play, if you capture a piece you do not know what piece you captured.5 I assume this would have a big impact on the performance of Dr. Ciancarini’s Kriegspiel program, but I could be wrong.
As a quick aside, I have reached out to Dr. Ciancairni to discuss his research, but I haven’t gotten a response. If anyone happens to know him or anyone else at the University of Bologna, let me know (not likely, but I thought I’d throw it out there haha!)
This is the power of information in Kriegspiel and in warfare. The best AI in the world will underperform if it cannot gather accurate data about its environment.
Deception
The rapid development of AI has led to huge breakthroughs and innovations in a variety of fields. But blind faith in the inevitable advancement of AI on all fronts is, in my opinion, unwarranted.
This is why, as I argued in the last essay, humans and AI will need to work together in the future. AI is excellent at analyzing data that it is given and even filling in gaps, but humans have evolved to survive in uncertainty. The more that sensors can remove uncertainty, the better that AI will be able to perform and the more advantage it will convey to its human masters. AI by itself, in the abstract, is not enough. We must focus not only on feeding our AI-powered systems good data but also actively working to feed our opponents’ systems bad data.
In the Blind AI section, I made the point that a strong human that can see all the pieces will always beat the best AI that could ever be developed that cannot see all the pieces. But imagine how much more dominant the human would be if the computer was given false information.
As Sun Tzu famously said, “Logistics is mid, but deception is FIRE!”
Thanks for reading The Distro!
Substack has introduced a new messaging feature, so please send me a message and tell me what you thought about this post!
Further reading:
13 parigi.ppt (unibo.it)
Monte Carlo tree search in Kriegspiel (uci.edu)
I have been playing chess and Kriegspiel my whole life. I am an average player. I know how strong Magnus is.
This is also why Kriegspiel is so different from regular chess. You have to use your pawns to screen for your pieces and conduct reconnaissance. Regular chess openings go completely out the window when pawns can sneak up and snatch a knight, as they did in this case.
I grew up calling these the “Cincinnati Rules.” The way I play, the referee always announces legal pawn captures.
FWIW, I think you're 100% on the right track here. The game of Diplomacy is another where AI researchers have been trying to do clever things with some success despite the challenges of dealing with almost purely subjective information about player intentions.
But we're all just wet robots, deep down, so an AI can pick up patterns. Oddly enough being consistent and honest in a game where most people play to betray is a strategically superior choice on the whole.
Evolution did a lot of handy programming work to make biological minds extremely good at handling ambiguity compared to any digitial system yet devised. My personal theory is that to make AI that could replace humans in a general sense you'd have to make AI cognition mimic human... which invariably would result in robots that act about like Bender from Futurama.
What do you do when your fancy warbot starts to question the meaning of their existence?
But the real question is, how do we challenge you to a game of Kriegspiel!?