Brain or Mentor, Classroom or Arena: Espousing the Role of M&S in Software Development for Unmanned Systems

by Dale Johnson

U.S. Air Force Photo by SrA Joshua Hoskins

U.S. Air Force Photo by SrA Joshua Hoskins

Much has been speculated about the future of aerial combat and the role that unmanned systems could play in it, but clearly much work is left to do to create the “digital pilots” of those future systems. There’s also the related challenge of knowing how to “train” and develop these digital pilots to maximize their capability and utility. Fortunately, modeling and simulation (M&S) software already offers both the core logic for future digital aviators, as well as the ideal virtual environment to run them through their paces to develop their tactics, techniques, and procedures (TTPs).

The U.S. Air Force has adopted the name Collaborative Combat Aircraft (CCA) for its future adjuncts to manned fighter aircraft, and for CCA to be a benefit to their flight mates, they will have to have enough autonomy so that their every move doesn’t have to be controlled by a remote operator (as is currently the case with Predator/Reaper drones). Initially, this autonomy may be limited to actions such as “fly this heading until you are told where/whom to shoot” or, for two sensor-equipped CCA, “fly these offset headings until you can triangulate the full enemy presentation.” But those roles are likely to evolve as human pilots grow accustomed to their electronic counterparts.

So where do we recruit these digital pilots? Rather than having to write the requisite software from scratch, the fastest route would, in general, be to find similar code and rehost it for one’s use. Fortunately, digital pilots have existed in the M&S world for decades. For most manned simulators, not all the cockpits are manned, with many of the players being digital (as in constructive simulations).

Brawler Crest

In the case of the air-to-air engagement model Brawler (the code this author is most familiar with), which started back in 1978, all of the “pilots” are digital; and they make all of the cockpit decisions. More recently developed engagement models, such as the Advanced Framework for Simulation Integration and Modeling (AFSIM), use the modern logic of state machines and decision trees to set up the complicated behavior to be simulated.

Any or all of these models or simulations could be distilled down to create the core “brain” of our new CCA pilot. A basic autopilot can already handle commands such as “steer to heading” or “climb to and maintain XX feet,” but these M&S-based pilot models should be better equipped to handle more tactically significant flight postures, such as “drag at heading YY until missile timeout,” or “turn into the beam for ZZ seconds then turn back into the fight” based on certain criteria being met.

Additionally, rather than spending thousands of hours debriefing pilots on best practices and TTPs, and then having to convert the collected data all over to flight rules and the biasing factors that will need to be evaluated to step through the decision trees, a further advantage of digital pilots is that they can benefit from machine learning (ML). With ML, the system steps thorough a myriad of tactics and alternatives until it finds the “best” choices for any given situation, and it banks those choices for future use.

This has been the case with the Defense Advanced Research Projects Agency’s (DARPA’s) Air Combat Evolution (ACE) program, where DARPA set up a digital arena for within-visual-range (WVR) air combat and invited contractors to design digital pilots to compete in a head-to-head tournament. At the end of the tournament, the winner was estimated to have undergone approximately 20,000 hours of WVR air-to-air combat prior to entering the competition. That level of training is simply not possible for human pilots.

Admittedly, once the ML has “taught” the CCA the “best” tactics, human pilots will likely have to comb through all those alternatives and weed out the implausible. Most survivability practitioners are all too familiar with the holes in M&S logic exploited by the system to select technologies or tactics that are not physically or practically possible. As the saying goes, “all models are wrong, some are useful.” However, a CCA that may have capabilities beyond that of its human counterpart (e.g., 15 g’s in any maneuver direction, not just +9/-1.5) may evolve to tactics that mortal pilots have never even contemplated.

Another area in which artificial intelligence (AI) and ML excel is pattern recognition. Because the survivability community and others have been recording Red Flag and other exercises (e.g., Northern Edge, Maple Flag, Green Flag, etc.) for many years, there should be ample libraries of telemetry and radar data for the AI/ML to pore through and distill current tactics as well as potentially exploitable vulnerabilities. All that is then needed is a venue or “digital arena” in which to experiment.

And where can we find these digital arenas? Once again, the M&S community is already surrounded by them. Large, manned simulations such as the Virtual Warfare Center (VWC) and Joint Simulation Environment (JSE) might be well served if, for a certain percentage of their runs, the digital pilots were allowed to innovate and evolve, learning as their human counterparts reacted to changing tactics. Understandably, these opportunities would probably be limited, as these expensive resources would likely be constrained to red air “sticking to the script.” However, perhaps some accommodation could be worked in on the margins and possibly take advantage of other downtimes in the schedule.

AFSIM LogoThe truly wide-open arena, where potentially millions of alternatives and counter-reactions could be played out, is in the all-digital simulations such as Brawler, AFSIM, the Joint Anti-Air Model (JAAM), and many others (the so-called deep-learning phase of ML). This is the classroom where the “brain,” which may have actually originated in one of these models, gets to try things out; where it may be “mentored” by the more established digital mental models; and where it can see how its digital adversaries react.

In conclusion, thousands of years of combined experience in both actual and simulated air combat have been distilled into the many models and simulations employed across the Department of Defense and industry. It would thus seem a shame not to draw on that experience when programming our initial CCA platforms. Furthermore, if we want to give future CCA all the advantages we can possibly give them, our M&S systems arguably offer the best starting point for creating a virtual “shoot house” in which those aircraft can train and maximize their capabilities and effectiveness.

About the Author

Mr. Dale Johnson is a recently retired member of the Air Force Secretariat’s Office of Studies and Analysis (SAF/SAFM) and has served as the Brawler Model Manager since 2004. In addition to his previous 20 years of active-duty service in the U.S. Air Force, he served as a Government civilian from 2009 until 2024. Mr. Johnson holds a B.S. and M.S. in aerospace engineering from the University of Michigan