The big Air Force news this morning is that an Artificial Intelligence algorithm easily defeated a human F-16 pilot in simulated aerial dogfighting in a 5-0 sweep. The three-day AlphaDogfight trials were hosted by DARPA.
I am not particularly surprised. I have never been a fighter pilot, but dogfighting–like many aspects of flying–strikes me as an exercise in precise energy management and control systems theory; the precise and timely application of proper control inputs will conserve energy and maximize the probability of ending up at the right time place at the right time. That is the kind of bounded task at which machine learning should excel (managing the complex, ambiguous, and open-ended environment in which dogfighting might occur is another story).
However, I have always worried about the ability of the DoD bureaucracy to manage AI algorithms. Our acquisition processes were designed for hardware; we spend years building, testing, certifying, and fielding a widget. Then we spend years sustaining it while we build the next version.
Software completely upended this paradigm, because in the modern world, software is released into production in the earliest stages of the lifecycle and new changes are released into production continuously–sometimes hundreds of times per day. Software obliterates the distinction between Research and Development (R&D) and Operations & Maintenance (O&M). It compresses requirements generation, design, building, testing, evaluation, and fielding, and sustainment into a tight circle rather than a linear process.
DoD has had severe struggles adapting to the world of software. A tireless assault by insurgent software factories–like Kessel Run, Space CAMP, the Defense Digital Service, Kobayashi Maru, Tron, BESPIN, and my own Rogue Squadron–has begun transforming the system. These organizations are making progress but it’s a bit like trying to steer the Titanic, and we all have scars. We certainly aren’t there yet; a lot of “agile” is still just agile BS. DoD rarely practices Continuous Integration and Delivery, and many DoD systems are still updated only intermittently–after new approvals processes–by handy-carrying memory sticks or even 1970s floppy disks.
AI algorithms strike me as an even harder challenge, because algorithms that continuously learn could literally change hundreds of time per second. If large parts of DoD still require lengthy approvals processes for new software versions, how on earth will we manage constantly evolving algorithms? Especially given how opaque and mysterious many algorithms are? One of the fundamental challenges with evaluating machine learning algorithms is that they are rarely explicable; results are achieved by manipulating parameter weights in models. They rarely rely on logic, heuristics, or features that can be articulated to mere mortals. They also have bizarre failure modes: they can be brittle, overfit to particular training sets, or trained on data that does not accurately represent the real world. This means that algorithms require human oversight, but DoD lacks the knowledge, skills, tools, and processes to manage AI effectively.
I am so concerned about this that I wrote a short story a few years ago, specifically to help DoD leaders understand the issues in play. In Fitness Function, bureaucratic inefficiencies throttle the vast potential of AI–and ensure overmatch by a near-peer competitor. If you are interested, you can read it online at CIMSEC or download a free copy.
Image courtesy of Breaking Defense