"Krasniy Oktyabr" Behavior Demonstration
The Scenario
The times are difficult. Once, you were a submarine commander in the world’s largest navy. Now your employers and crew have changed.
You are in command of the ex-USSR Typhoon class nuclear submarine Krasniy Oktyabr (also known as Red October) bought on the black market. You have demanded a payment of 10 tons of pure gold from the state of Colorado, as your pirate crew believes it is full of gold mines. (Okay, so you would not do such a thing.) Colorado has disregarded all your cellular phone calls, so now you must launch your single missile to show that your demands are in earnest. The missile has 10 MIRV (Multiple Independently-launched Reentry Vehicles) warheads and over eighty very clever decoys, which look like another warhead to the Colorado radars. You will show them! You have a probably reliable map of various strategic targets in Colorado, and you and the crew decide to hit these. Let’s see if Colorado’s defenses (directed by LINIAC™ neural net decision structures) are capable of deflecting your nuclear ire!
The Challenge
Can you get at least one of your warheads to hit anything, anywhere, in Colorado? If you can, you can really show them what a dangerous opponent you are! (But will they pay?) Even second best counts here for something. Colorado has 240 ABMs (Anti-Ballistic Missiles) to start with. It really makes no difference if your warheads don’t get through. If, after your attack, Colorado has only 130 or fewer ABMs left, they can’t resist a second attack, and they know it! Moreover, they don’t know that you had only one missile, so they will probably pay up! Even if they have more than 130 ABMs left, they still may well pay. Let’s see if you can outfox them, forcing them to expend a lot of ABMs against your attack! There are two factors that you can more or less control: targeting and launch locations. The computer will grade each attack:
Result
At least one hit
Colorado has 130 or fewer ABMs left
Colorado has 131 to 140 ABMs left
Colorado has 141 to 150 ABMs left
Colorado has 151 to 160 ABMs left
Colorado has more than 160 ABMs left
Grade
A+
A
B
C
D
F
Of course, luck has something to do with this. The five sensor systems of Colorado may or may not work. You will only know this after you have launched. If you are lucky, their radar will go down, and one or two other systems may also go down, just as you attack. Their ABMs, on average, have only an 85% to 88% kill probability if they can intercept your warheads (and they can only intercept at relatively shallow angles). They can only launch in very limited launch windows (see Defenses).
Targeting
When you start, you will see a map of Colorado on the left, and the rose-colored Attack Planning Panel on the right.
The software used by Krasniy Oktyabr is pretty clever; it will select 10 key targets on its own and will also generate decoy attacks. Some of the decoy attacks may just saturate your targets, but other groups of decoys will pretend to attack key targets you are not really attacking! This is to mislead the defenders and force them to expend their ABMs against mere decoys. When your targeting is in the “Auto” mode (selected by the radio button on the top part of the Attack Planning Panel), the default targeting mode, the computer will do all targeting for you. If you don’t like the computer’s selections, you can click on the “New Targeting” button, and you will have a whole new targeting configuration. The list of the ten targets is shown in the “Warhead Aim Points” list.
You don’t have to leave targeting to the computer. You may pick targets on your own for your ten warheads. Just click on the “Manual” radio button, and it will clear the Warhead Aim Points list. Now you can pick your targets on the map with the mouse. You can develop your own strategies to fool the defenders! The software will still program your 84 decoys; it will support your attack plan. You don’t have to select all 10 targets manually; those warheads not assigned targets on your list will be assigned targets automatically by the software.
The map display on the left side shows your targeting. But don’t be satisfied just with your targeting; check out the approach paths by selecting “Approach Info” on the bottom of the Attack Planning panel. If you think you should improve the approach paths, select a different launch location! Just click on the “Set Launch Point” radio button.
You can start your attack at any time by pushing the red LAUNCH button. (Notice some of the old label of the original owners is still there.)
Launch Locations
One way to improve your chances is to change your launch locations. All of them are in the seas around the former USSR. (Your current masters felt it was best to lurk around there, so they will not be suspected!)
There are six launch locations. You can change these by opening up the Launch Locations Map (select the radio button “Set Launch Point”). As long as on the right you see the pink Attack Display panel (that is, before you launch), you can select a new location by clicking on the “Push to Change Launch Location” button. You will get the shortest flight times from locations near Kamchatka. On the other hand, you may devise strategies that exploit launches from the Arctic Ocean, or from the Poliyarniy area above the Kara Peninsula.
You can check your approach paths on the Approach Map (select “Approach Info” on the Attack Planning Panel).
Push the red “LAUNCH” button to start your attack!
Defenses
What does Colorado have to defend itself against your attack? Your spies have obtained the following classified intelligence information:
Defense Assets
Short-range ABMs
Mid-range ABMs
Sensor System
Virtual Lieutenants
Reliability
85%-88%
85%-88%
work 90% of the time
work all the time
Quantity
120 total in 6 batteries
120 total in 6 batteries
5 systems
200
The six mid-course ABM batteries are located outside Colorado (which can be seen on the Approach Info display). Each of these has a load of 20 ABMs. They have a 600 nautical mile maximum range.
Five of the six terminal defense ABM batteries are in Colorado, and the sixth is just a few miles north of the Colorado border, in Wyoming. Each terminal defense battery has an initial load of 20 ABMs, with a 150 nautical mile maximum range.
The ABMs can intercept your incoming warheads (or decoys, or debris) only at a shallow angle (your spies did not get the exact information on this). They must launch early, because their speed is only about 1/3 that of your warheads. Therefore, one factor in your favor is that these ABMs have a small launch window, and they must be launched when the defenders don’t have much information. Even if the ABMs intercept your warheads, there is only an 85% to 88% change they can destroy them.
The five sensor systems are not very reliable. There is about an even chance that at least one system will not work. Often, two systems will fail—or even three. With some luck on your side, maybe only two systems will work. Each of these systems is different; some predict the impact point, others try to recognize the shape of the incoming objects, and so on. They are not very good at the beginning of the attack, and most are just as likely to give wrong information as right for the first 25% or 30% of your attack. Later, they get better; but it is only in the last two minutes, when atmospheric re-entry clearly separates warheads from decoys and debris, that they can tell for certain what is coming.
The “Virtual Lieutenants” are tiny, but very fast software modules. They try to determine from unreliable (and often missing) sensor data what each “track” contains—is it a warhead or only an inert object? They also decide how many ABMs should be fired against each track and when they should be fired. There are two hundred of these “virtual lieutenants” “manning the desks” in the Virtual Defense Command Center (see more in the Virtual Lieutenants section). Their chief advantage is that they are fearless, quick, and numerous.
Execution of the Attack
You have launched your attack! This will now go very quickly. The simulation runs 60 times faster than real time; one minute of your warheads’ flight time (which can total anywhere between 19 and 25 minutes) takes one second in the simulation, except on a very slow computer.
The top of the area on the right is the light blue Defense Information Panel. About the only control you have now is PAUSE and RESTART, in case things seem to go too fast! The panel shows information as seen by the defenders in the blue area: where the attack seems to be coming from, the number of incoming tracks detected, estimated time until impact, and the number of ABMs left. This is shown in the uppermost light blue area.
The current track classification is shown below the Defense Information Panel. At the beginning, everything is a threat (a “Bogey”), but the “Virtual Lieutenants” in the command center pretty soon start classifying some tracks as confirmed inert. Inert tracks are mostly decoys, and some of them are pieces of debris from the warhead shrouds and supporting structure. You will also see the health of the defense’s sensors. “5/5” means all 5 systems are operational. You, of course, hope for 0/5, while they would like to have 5/5. However, this is neither under your control nor theirs. It is pure chance.
Critical events are shown at the very bottom of the blue area. A critical event happens when one of your warheads hits!
Ground truth is shown in the grey area. You will see that many of the ABM launches are committed against targets that in truth are inert. The defenders will do this because they have not enough information and not enough time left!
Why doesn’t the defense just shoot at all targets? They have 240 ABMs against 100 incoming objects. The problem is that not all ABMs can reach all tracks, so the defenders really don’t have enough ABMs. In addition, they must balance two factors: preserve as many ABMs as they can against another attack and make sure that all real bogeys are killed. That is the challenge of the defenders.
The map on the left shows how the attack is progressing. Black squares show bogeys perceived by the defense. Yellow squares indicate the flight of an ABM against a perceived bogey. When a perceived bogey is killed or confirmed inert, the display for that track disappears.
After the attack is over, you will receive your score for this attack (see The Challenge). You should also review what really happened by clicking on the After-Action Review button.
After-Action Review
The After-Action Review display can show you what happened. Click on the After-Action Review button after your attack is completed. The first chart you see shows the Combat Data.
The Combat Data Chart shows the time history of four items of information. The blue bars show the number of mid-range ABMs launched, grouped for each minute. The magenta bars show the number of terminal defense ABMs launched.
The solid red line shows the number of actual warheads in the attack, and the red dots show how many of these warheads are alive at the end of each minute of the attack. You will notice that there is a considerable time delay between mid-range ABM launches and the destruction of the warheads. These ABMs are shot into the path of the very distant, but very quickly approaching, warheads (coming at about 4 miles per second—over 6 km per second!). The delay between terminal defense ABM shots and destruction of their targets is considerably shorter.
The “Show Detection Data” button shows how much the defense knew about the incoming tracks. When all five sensor systems are operational, about half of the inert tracks are identified very early. When some key sensors are not working, the defense often must assume for about half of the flight time of the warheads that all tracks are bogeys! You may switch back and forth between these charts, and you may also elect to look at the map (which will be empty unless one of your warheads hit Colorado). You may start a new scenario; just click on the “Start New Scenario” button. Show that you can outfox these virtual lieutenants!
The Virtual Lieutenants
Who are these fearless, quick-thinking virtual “individuals”? They do their job a thousand times faster than you can blink an eye, and they don’t blink an eye even when a hundred bogeys seem to attack their headquarters, and all sensor systems are down!
On the positive side, we believe that they are the first generation of 21st century Artificial Intelligence—that is, software that tries to do those things that humans do well, but computers have great problems doing. We are talking about “judgment”, “emotion”, and “behavior”—all those things that cannot be easily emulated in algorithms, rules, or dazzling graphics. They are (sort of) based on the “neural net” paradigm, which is really just massively parallel, distributed processing. We say “sort of”, because we think we have made a “giant leap” ahead of “classical” neural net research, focused very strongly on practical application.
Where these “lieutenants” operating in our “Virtual Defense Command Center” differ from the usual approach to “neural nets” is that they emulate human organizations. Therefore, the usual research question “How many layers do you have in your model?” is wholly meaningless in our approach. Each “desk” in this command center has only one “processing layer”. On the other hand, how the layers interact with each other, which layer gets its input from which other layer, is decided “ad hoc”. There is no need for one “layer” to strictly follow another in a fixed order, and there is no need for one “layer” to get all its data from another “layer”.
These “virtual lieutenants” are specialists. They become specialists very quickly, but, of course, it took time to determine what factors they should consider. They can consider many factors simultaneously, as they are doing in this demonstration. They interoperate with each other as necessary, emulating very much a human organization.
The Technology
The Virtual Defense Command Center directing Colorado’s defenses has 100 “decision structures”, built (instantiated) from a single Java class file (LINIAC.class). This file is small; the size is a mere 1430 bytes. It contains all the methods that are needed to perform the “thinking” needed by the “Virtual Lieutenants”. They “think” very quickly; each “lieutenant” will make a decision in a few tens of microseconds on a 233 MHz Pentium running under an Internet Explorer 4.0 browser.
Each “lieutenant” is “implemented” as a LINIAC class artificial neural net. Two of these “lieutenants” work together to form a “decision structure”. You might think of the decision structure as a small team whose members are working together, following a “standard operating procedure”. Each “virtual lieutenant” is either a detection specialist (correlates all the instrument information and other data), or a targeting and launch control specialist. These small “virtual teams” work together with all the other “virtual teams” to determine which tracks must be attacked and when.
The LINIAC class is instantiated only twice: once to represent the “learning” of the detection specialists and once to represent the learning of the targeting and launch control specialists.
These “specialists” work so fast, that on a single-processor, 233 MHz Pentium PC, the total work of these 100 “virtual teams” takes up only about 1% of the processing time—and the scenario is run at 60 times faster than real time! Most of the processing time is spend in doing the mechanical and graphics aspects of the simulation. The “Virtual Defense Command Center” uses negligible computing resources, either in processing time or in memory.
The “virtual specialists” have “learned their trade” from a retired Army Colonel, who spent altogether about 2 hours “training” them. The targeting and launch control specialists have learned their trade by “seeing” 14 examples of the “right kind of behavior”, and the detection specialists have “seen” 23 examples of the right kind of behavior. From these learning sessions, they have both learned enough to deal with thousands of different combinations of factors they may have to face.
What does one need to know about “neural nets” to train one of these LINIACs? Absolutely nothing! The only “computer literacy” needed is that the trainer must be able to click on a mouse! What the trainer needs to know is, of course, what is being “taught”: the LINIACs will only emulate the trainer and generalize “the kind of thing they see”.
How fast do they “learn”? Each example is absorbed normally in a fraction of a second. Most of the time in training is spent by the trainer thinking up examples, evaluating the LINIAC’s “behavior” through examples of situations it has never seen, and clicking on the mouse.
Credits
The concept of “virtual organizations” has been conceived and developed at Pathfinder Systems, Inc. We own all copyrights and all rights to the underlying software.
None of this, however would have been conceived and developed without other people and organizations.
General credit is due to Professor Marvin Minsky of MIT, who, in the early 1960’s, inspired the author of these concepts with his thought-provoking lectures, and to whom is due the credit for proving that linear neural nets “ain’t much”.
Commander Denny McBride of the U.S. Navy was responsible for organizing the first conference on “Behavioral Emulation of Computer-generated Forces” in the fall of 1990. The author’s direct inspiration for the methodology came from informal discussions at that conference.
“Parallel Distributed Processing” (MIT Press, 1986) by James L. McClelland and David E. Rumelhart provided many of the examples, and the inspiration of extending the Interactive Activation and Competition model for a practical methodology (the model from which the IAC in LINIAC comes).
The Ballistic Missile Defense Office, under the supervision of Dr. Claire McCullough of the U.S. Army Space and Strategic Defense Command provided the funding in a small SBIR project for building the predecessor of this “Krasniy Oktyabr” demonstration. Colonel Robert T. Reed, USA, (Ret.), a co-owner of Pathfinder Systems, has trained the “virtual lieutenants”.
The U.S. Army Simulation, Training, and Instrumentation Command (STRICOM) provided initial funding for more detailed research to start exploring how the large number of “role-players” in military exercises could use the techniques described as “virtual headquarters”.
We thank all our friends without whom this work could never have been done.