20 June 2018

Machine Strategists & the Future of Military Operations

By Thomas Keelan

What do drunk Google searches and war have in common? They’re both chaotic, incoherent and occasionally regrettable. They are also both more effectively solved by machine learning. Modern technology has made military strategy more complicated than ever – just look at China’s recent installation of cruise missiles on its artificial islands in the South China Sea. It’s becoming harder and harder for humans to keep up. Far away from the “tactical edge” of soldiers and weapons, smart algorithms like Google’s RankBrain will soon be needed to analyze the mountains of data and invent new military strategies. 


There are multiple approaches to creating machine intelligence (clustering, gradient descent, random decision forests), but in basic terms, it’s done by combining complex equations to function as a sort of artificial brain. Through intricate mathematical modeling, these algorithmic “brains” are trained to identify patterns in existing data. They are then asked to make independent decisions about new data – to learn by themselves.

RankBrain is a machine learning system that Google added to its pre-existing search algorithms in 2015. It was introduced to help Google searches work “semantically” – that is, finding answers not to what you typed, but to what you meant. Google’s other algorithms already did this to some degree, but 15% of user searches are completely newr4 and are often complicated, ambiguous or downright indecipherable.

That’s where machine learning comes in. It uses patterns in language structure (the training data) to teach itself how these confusing searches (the new data) should be analyzed and sorted. In effect, it acts as an instant interpreter between the user and Google’s general search engine. RankBrain’s uncanny ability to find exactly what we’re looking for is a fitting example of machine learning’s inhuman speed, consistency and creativity. 

Such qualities will be critically important for future militaries. As the recent U.S. Third Offset Strategy demonstrates, modern militaries are being forced to adapt to the digital, data-centric age. Operations are speeding up, but there is now too much data being generated for humans to process. Machine learning can bridge this gap, and provide the countries that effectively invest in it with a military cutting edge. China and America are now doing just that – pouring billions of dollars into military machine learning projects.

So how exactly might these “machine strategists” help solve problems like the South China Sea?

Machine learning is still some way off from being able to usefully analyze fields as complex as military operations. An early hurdle is providing such algorithms with adequate training data. By one estimate, they need at least 10 million labeled examples to exceed human performance. For Google this is easy – with three billion searches processed every day, they have one of the largest data sets in the world.

To approach this quality of training data, a machine strategist would need to be fed information on two levels. The first would teach it the structures and logic of military operations, just like real search algorithms are first taught the structures of language. It could be given summaries of every battle and campaign ever fought. These could be broken down by a number of combatants, maneuvers, terrain, technology level, tactics used, and countless other categories.

Second, a machine strategist would need to add real-world context to the abstract contours of warfare. This would work in the same way as Google using the trillions of old searches to bridge the gap between formal language and the rushed, stream-of-consciousness typing of users. To do so, a machine strategist would be given every scrap of data on each nation’s current resources: armed forces, population, economy, infrastructure, even culture. 

Across this dizzying depth and breadth of data, the algorithm could rapidly identify patterns and opportunities concealed from humans.

Such encyclopedic strategic knowledge (the training data) would enable a machine strategist to teach itself to analyze novel conflict situations (the new data). It would then interpret the chaos of a future war into a clear strategy for human commanders. Moreover, as with mangled Google searches, it could creatively adapt to unforeseen outliers – like biological or nuclear attacks – and provide optimal solutions.

One key benefit of these algorithms would be an ability to prepare as well as react. A machine strategist could preemptively run countless simulations of different types of conflicts, to inform leaders of their best options and highlight possible areas of vulnerability. It might also be able to offer suggestions on how best to prepare a nation for war during peacetime, coordinating all resources (human, industrial, economic) for maximum effect.

Machine learning is coming soon to a conflict near you. When it does, its real impact will not be felt in merciless Terminator robots, but in unseen learning algorithms that make warfare more systematic, creative, and perhaps more lethal than ever before.

Thomas Keelan is a research assistant at Hudson Institute.

No comments: