AI is engaging, Sir!
Víctor Manuel Sobrino García
Teniente Coronel. Piloto de combate de F-18 en el Ejército del Aire| NGWS/FCAS Spanish NPO deputy chief | Grado en Ingenieria Informática. MSc en IA avanzada | Investigador Doctoral en IA en CNIO.|MBA
Today I have found myself watching a 5 hours video on YouTube about one of the latest advances on AI application to the military field. Heron, a company that is dedicated to build train and test AIs has developed a model that is able to dogfight against a human pilot and win in less than five minutes (what is too much when talking about dogfight) five times in a row. Moreover, the human pilot never get to win a single engagement. Fly, Fight, Win!
Ok. Thats nice, pretty impressive, but is just an application of the game theories that does not differs a lot with any other game. The main differences is that the pieces in this board are a little more expensive. Let me please explain myself.
For those that are not familiarized with air combat maneuvering (ACM), a dogfight can easily be projected in their mind as an intuitive game where a single sparkle of side thinking combined with an outstanding aircraft handling capabilities will always make you a winner. "I am gonna hit the brakes so he will right fly by" Maverick said. Real life dogfighting is no so epic. Basic Fighter Maneuvering (BFM) deals with the management of the availability of energy on your jet. It deals with the relationship between kinetic and potential energy that are provided by the speed and altitude that is generated by the appropriate use of the thrust, load factor, turn radius and angle of attack (AOA). In fact, the Energy vs maneuverability diagram depicts what are the best opportunities for your aircraft in a given energetic situation.
In this kind of diagrams what you can see is the speed against turn rate. It is obvious, if you don't have speed you won't be able to pull a single G, that is, to turn "I feel the need... the need for speed" (This one is true, at least, Mav), but there is much more information enclosed here. In the EM diagram one can find what is the maximum number of Gs that the pilot can apply to the jet depending on the speed and altitude, as well as what is the turn radious they will get whit a given number of Gsm etc, so... that's it, all the information is here, and thanks to those crazy engineers that dedicated their life to make us, the pilots, fly, we have it.
So... What I want this for.
Well I can't briefly reduce 15 years of dogfighting in some sentences, nevertheless when talking about the basics, in a close combat, the one that is "inside the circle" wins. That is (more or less) all.
When two jets get to a neutral engagement (head on) what each pilot wants is to put the oponnet in front of the nose of their plane (ok, lets forget HMS, HMD, Sidewinder X or IRIS-T this time, ok?). Why? To shoot. To do so you had to turn against the other′s tail. (Easy, if you want to go behind the opponent plane you have to point to..... yes! the rear part of the opposer plane). So, once engaged, both jets will start to turn (just one note, depending on the maneuvering category of the jets, maybe you will like to search the nose of the opponent instead of the tail, but, both jets will be turning making circles anyway) making circles in the air. Now imagine that you are driving a car in a roundabout, and you want to overtake another car, what you have to do is to go into the inner part of the roundabout and even without accelerating you will make it. Why? ok, I know that is common knowledge, nevertheless..... because the track is shorter. But, why the track is shorter? Because the radius of the circle you have described with your car is smaller (2*pi* radius), so, what you want is this case is to reduce your turn radius.
This situation represents the basic of dogfighting. If both jets are turning with the same turn radius, they fill face off every lap. To avoid this, one jet must make the track sorter, If the track is sorter, then, this plane will get to the next crossing point sooner. If this is the case, what this pilot will find is that the plane will cross with the other jet with a certain angle, there won't be a head-on pass anymore. If this process is repeated lap after lap, in a given number of laps you will find enough angle to attack and, hence, to shoot. This is called the decoupled circles theory
I know the basics, now what?
Well, once this is known you can easily understand that the difference here relies in knowing that capabilities of your jet when your energy is know (speed, altitude, load factor...). For instance, some jets have high or none AOA limitation (like the F-18). Since AOA is the angle formed by the air and the wind, if you have enough AOA you can make your jet drift away from the trajectory to head you opponent and shoot in a kill action to, later, quickly return to a low AOA situation (because to pull with high AOA kills your energy). To do so the pilot have to be aware of in what zone of the EM Diagram is the jet flying.
The same happens with your opponent (Because there is another guy here).
These are basically the tools, a jet, and physics. This is it. Well and a lot of data. But.. AIs are good when dealing with data, aren't they?.
I am still awake! tell me about AI.
If you haven't felt sleep yet, this is the part dealing with AI. Once you know this, we are facing is , hence, just a game. We have a fitness function (a goal) that is the relationship against my and my opponent jets energy and relative positioning. Moreover, I have a diagram and some Newton′s theory equations that can help me to build an heuristic that will guide my AI through the decision process. So, what we can do (for instance) is to train the machine to identify certain situations and to act accordingly to the data available, to compute the data and to perform a typical decision process to learn what to do. Moreover we can train it with SOAR, or with just other technics (let say Evolutionary Computation) to, once a situation is identified provide the best solution in the solutions space. This will ease the computational load in real time. There are a lot of techniques you can apply..... or not.
There is something we are missing here, is something important. And this is part of the dogfight magic. This is the part the human does, this is why Maverick (and Ice Man, that is the really cool guy of the film BTW) aways win and cougar had to quit. There is an opposer in the game and you don't know a single thing about the situation of the opponent plane.
When playing a game there is some information that is missing and some other that is known. A player usually knows all the information of the game itself, (i.g) the position of the pieces and the allowed movements in a chess, go! or any other board game. What the player has to do is to figure out the next movement of the opposer (the tactic), but this is not the case of the dogfight. As dogfights evolves very fast, the jet can easily and quickly move from a position in the EM diagram to another. And you know what?, the enemy won't tell you "Ok! Now I am at 15 thousands with 350Kts and 35AOA pulling at 4 gs" you have to made it up. How? Based on you experience and some indicators.
And....Here is the rabbit out of the hat
When I watched the YouTube presentation today I was immediately thinking on this people that is right know building their bunkers and filling them with water and food to prevent the effects of the next skynet actions. Of course AI is great and powerful, but there is still a track to walk. Maturation of the technology (TRL) is still low. And magic is full of ticks.
The dogfight game is easy when you know both your and your opponent conditions. Moreover, if you have this information, an enhanced computation capabilities together with a precise control surfaces handling capabilities, it will be difficult yo beat you. Impossible I would say, It will only depend on you the capabilities of you jet (if it is an F-18 you win, that is it).
What I will like to know, now, is what is the data that was used to feed the AI. Did the AI take the speed, altitude, AOA, etc, information of the opponent as an input and use it to perform an estimation of the "nest move" or did it just watched the opponent plane in the screen and tried to "guess" what is the energy of the other plane and act in accordance to this supposition. I personally think that is the first one, if this is the case, and you are still with me, you will notice that, it is not so difficult then. Don't miss understand me. anyway all this is impressive. To develop this AI requires a lot of BFN-ACM study, computation, and training, they have to teach it that crashing or been shot down is bad! that it had to win!. There is a lot of good work there, but, the magic is in the guessing part. This is what will make the AI reliable to use it for real time purposes or not.
This part is difficult, very difficult, because there is a lot of "under the table" game here. Sometimes in the dogfight, things are not like they looks like. A fighter pilot has to understand what is the energetic condition of the proponent by watching the relative position and motion, the nose elevation and the track the plane is following, and all this factors can be easily modified by the opponent to make you fall in the trap. If the AI is able to understand this and win in this condition, if is the case, then it is more than more awesome. It will in fact be a game changer.
One more thing... sensing
Interesting? I hope so, nevertheless here is where the theoretical approach lands in reality, how can we make our incredible AI to accurately guess the energetic conditions of the other jet? By adding good sensing capabilities to it.
But this is another issue
Thanks
Especialista Medicina Aeroespacial. Médico Examinador Aéreo. CIMA. Ejército del Aire y del Espacio
4 年Enterito me lo he leído! Muy interesante! ????????????
ADMINISTRADOR AEROPLANTAS YPF - CóRDOBA
4 年Excelente analisis! Particularmente creo que la "simulación" llevada a cabo busca una aproximación en escala para futuros drones ESCORT y AUTóNOMOS. Y en el proceso (para la toma de decisiones) exigen el dise?o conceptual en el ambiente más exigente y dinámico, tal cual es el DOGFIGHT. Totalmente de acuerdo que el "juego" debe tomar datos propios y del oponente para poder acelerar su OODA LOOP y ganar...
Experimental Test Pilot at AIRBUS
4 年Muy buen artículo Sobrino. Completamente de acuerdo contigo. Creo que en el combate cerrado efectivamente entran en juego muchos más factores que los que hayan podido incluir en esos algoritmos, sin quitar el mérito que tiene el haber desarrollado esta magnífica aplicación. Un abrazo.
NATO
4 年Muy buen artículo Sobri.
ACAR Tablada (Seville) barracks Commander @ Spanish Air & Space Force | Master in HR Management
4 年Great article. Sort of a lesson to me. Thanks