Real-Time Adversial Intelligence and Decision-making

Moderator: jimothy_183

Post Reply
bria
Posts: 20
Joined: Sun Nov 22, 2015 11:01 pm
Location: France

Real-Time Adversial Intelligence and Decision-making

Post by bria » Tue Jan 26, 2016 10:55 pm

Hello,

From another subject, I've tumbled on a (old) darpa project called RAID (Real-Time Adversial Intelligence and Decision-making ) (if you search google, please don't use only RAID, as it's also a well known acronym in computer hardware).

The idea behind this project is to raise Situational Awareness(SA) by using (machine learning?) algorithms .
They get informations from sensors and other recoinnaissance , and give back Level 2 SA (comprehension) and Level 3 SA (prediction, pattern matching).

(At least, that what I understand)

Although theses tools doesn't have the intuition or the experience of a human, they seems to produce results with less error as soon as sufficient information is present. (One of the main explanation is the human cognitive bias, which help us greatly in speeding information analysis and limiting noise. This work by discarding information that doesn't conform to what we've got in our mind).

The PDF showing this subject :
http://www.dtic.mil/cgi-bin/GetTRDoc?Lo ... =ADA460912

(You could find the paper from which I'be heard of this program here : https://www.researchgate.net/publicatio ... _and_Cyber , but it's a little more specialised in the "cyber" aspect
Since they give the results analyses I've summaried above, it could still be an interesting reading.
)


Do you hear of similar program, or did you already use something similar to adapt your tactics and evolution as the situation change ?
Do you have any idea how well they could work in a noise/mis-information real-world ?

Even if it's more useful on a strategic view, do you think you could accept a machine helping you on theses aspects, instead of a flesh-and-bone help ?
Anyone to build an open-source machine learning tactical wargame algorithm to optimize tactics/....training :P ?

User avatar
Ryan
Posts: 2793
Joined: Sun Apr 10, 2011 3:10 am
Contact:

Re: Real-Time Adversial Intelligence and Decision-making

Post by Ryan » Wed Jan 27, 2016 2:08 am

Very cool bria. This is the first time I have ever heard of this.
bria wrote:Do you hear of similar program, or did you already use something similar to adapt your tactics and evolution as the situation change ?
Do you have any idea how well they could work in a noise/mis-information real-world ?

Even if it's more useful on a strategic view, do you think you could accept a machine helping you on theses aspects, instead of a flesh-and-bone help ?
Anyone to build an open-source machine learning tactical wargame algorithm to optimize tactics/....training :P ?
A similar program? No.
Similar to adapt tactics? Yes.


You can do something in real-time called an External or Internal Structure Analysis (ESA, ISA).

It is similar to a breachers assessment without the whole breachers log. It is done on the fly. You go up to the door, decide there and then.

Well, in this case. You are approaching the building - you can make fairly accurate guesstimates to the setup and layout of the building from outside and inside.

Okay so you see a small square shaped building. One opening in the center. A few windows.


What can you tell?


It is square shaped. About 20x20. You have some basic ideas in your head. It could be just a box room.
It could be split by a separating wall with or without a door, into two interconnecting rooms.
It could potentially be clusterfuck and surprise you with a hallway but that is unlikely for the space.
As you move up you can see more through windows, etc.

You're inside. You find an interconnecting room. Door closed.
This will more than likely lead into another box room.
Think of the possibilities of cupboards, wardrobes, a toilet which would create another room and angle to clear.

Get how it works?
You can progress this further with double doors = larger rooms. Vents = this or that, etc.

And if you have watched Moneyball. Hell yeah algorithms help. Most professional sports teams are doing this. An entry team is like a sports team.
CQB-TEAM Education and Motivation.

"Pragmatism over theory."
"Anyone with a weapon is just as deadly as the next person."
"Unopposed CQB is always a success, if you wanted you could moonwalk into the room holding a Pepsi."

User avatar
tacticalguy
Posts: 700
Joined: Thu Mar 29, 2012 4:48 pm
Location: Florida

Re: Real-Time Adversial Intelligence and Decision-making

Post by tacticalguy » Thu Jan 28, 2016 6:26 am

Nice summation of critical thinking, Ryan.
If you have `cleared' all the rooms and met no resistance, you and your entry team have probably kicked in the door of the wrong house.
(Murphy's Cop Laws)

The greatest enemy of a good plan is the dream of a perfect plan. (Von Clausewitz)

bria
Posts: 20
Joined: Sun Nov 22, 2015 11:01 pm
Location: France

Re: Real-Time Adversial Intelligence and Decision-making

Post by bria » Fri Jan 29, 2016 8:06 am

As I understand, ESA, ISA are used for static things (room placement and shape, architecture).

The "fun" point about RAID is to add some more degree to it with
- analysis of dynamic conditions (the enemy is moving from this building to this building , thermal signature or EM signature changed in the last hour)
- prevent any cognitive bias. The algorithm isn't enticed to one or another theory how the enemy must behave. He only used the fact to determines the outcome. ( -> and that's great for any training ! )
- prevent information overload and let the operator concentrate on his task.
- aggregate, in near real time, multiple feeds.




RAID is actually defined on a more wide area of operation than CQB, even if it was designed with MOUT in mind.
Could it be adapted to CQB ?


For example, with structural analysis in an assault, you must be dong it each time, alone. I presume you don't really have the time to get everyone opinion on what could be behind this door and how is shaped the room you will enter. Especially if one of your teammate had just been shouted from this door.

We could imagine that each of the teammate get camera which are are aggregated in a global feed then analysis in real time with their position. (Computer vision had done some big leap in the civil, and the detection of object and person could be pretty impressive ... and also pretty dumb if not done right with enough information.
For example With multiple feed of a same object, anybody could now reconstruct it in 3D. We could detect a face in a picture with a high probability, and even compare theses faces. ).

From this, it could pinpoint possible location of IED, threat, (machine gun nest, ...), establish an automatic plan and blueprint and redistribute theses information for all .
That for the static/architectural part.

Each time there a contact or a gunshot (or a simple detection), it will use it to augment his "theory" about how the enemy react.
It could then
* how and where the next contact will be the most probable (think of "hit and run" ennemy)
* Try to automatically classify the weapons.
* Giving the probable most safest way to exfil.
* avoiding blue on blue if there multiples assault teams.
* ...

Of course, in a live CQB, this would be pretty daunting to get enough information quickly enough for this to work. And then present it in a clear, simple way which let the operator do their jobs.

User avatar
Ryan
Posts: 2793
Joined: Sun Apr 10, 2011 3:10 am
Contact:

Re: Real-Time Adversial Intelligence and Decision-making

Post by Ryan » Fri Jan 29, 2016 8:28 am

:lol: Sounds too complex for me!
CQB-TEAM Education and Motivation.

"Pragmatism over theory."
"Anyone with a weapon is just as deadly as the next person."
"Unopposed CQB is always a success, if you wanted you could moonwalk into the room holding a Pepsi."

bria
Posts: 20
Joined: Sun Nov 22, 2015 11:01 pm
Location: France

Re: Real-Time Adversial Intelligence and Decision-making

Post by bria » Fri Jan 29, 2016 10:02 pm

Sorry, That's part of my domain (not on so complex design of course), so I'm bad at explaining it to others without technical details.
The idea behind is pretty basic : an extension of what you've already done. Only by a computer :)

Let's just defining a concept
We use an ISA/ESA approach not only for deciding of room's shape but also for where actually is the enemy and where he will be in 5-60 min. This is done by a computer.


Actually this only seems to work on larger scale battlefield (MOUT and above), since it depends on enough reconnaissances information, and to maintain reconnaissance throughout the process.


My supposition is : if we use each team member for gathering information (in a passive way : without interfering with their work) , could this be enough to apply the same system to the high intensity, low spatial dispersion, combat characteristics* of CQB ?

I Hope this explanation is clearer :)

* : this definition is mine. I've seen the posts on "pistol range or not pistol range" and so on.

User avatar
Ryan
Posts: 2793
Joined: Sun Apr 10, 2011 3:10 am
Contact:

Re: Real-Time Adversial Intelligence and Decision-making

Post by Ryan » Fri Jan 29, 2016 11:18 pm

If that were indeed possible, it would change the face of CQB forever. Especially templated hostage rescue.

I mean it would require a lot of things. You can use sensors to detect gunshots, for example. With distances. And triangulate to a given area.
That will triangulate to a location. That could be within a few meters accuracy depending on how they were set up, how open it was, etc.

That's one easy algorithm and start. They used similar in Iraq on Humvees. Real-time sound signature can detect location or, at minimum, direction.

This would probably require sensors to be set up at three to four separate points but that's possible if it is not an active shooter situation.

Analysis of building can be done with wall-radars. K9 with camera transmitted back to team is possible. Thermal signature too, as you say.
Thrown sensors could transmit messages when someone is near them. Microphones and cameras can be inserted through walls, under-doors.
You could then find the "hot zones" where the most activity is. Patrol paths for hostage takers. You could even note if a hostage taker pretends to be a hostage.

Prevent information overload could just simply be information through radio to say X is here. Or a handheld device to give direct information.
Or the sound devices could transmit to the operator. "Gunshot, right, 100m" or "Suspect 1 in Room 3, back right."

Videos on the back or front of operators can be done - the SASR/Commandos here in Australia have them on their chest rigs.
Am I on the right track here? You need something to give you information then use the algorithm. The more information you can get, the more accurate the algorithm?


:lol:
https://en.wikipedia.org/wiki/Sabermetrics
https://en.wikipedia.org/wiki/APBRmetrics
https://en.wikipedia.org/wiki/PECOTA

Okay now I've divulged into theoryville. :lol: I am really imagining this from the fixed hostage situation because it will give teams adequate time to set up and probably only one building with a few hostage takers to deal with.
CQB-TEAM Education and Motivation.

"Pragmatism over theory."
"Anyone with a weapon is just as deadly as the next person."
"Unopposed CQB is always a success, if you wanted you could moonwalk into the room holding a Pepsi."

bria
Posts: 20
Joined: Sun Nov 22, 2015 11:01 pm
Location: France

Re: Real-Time Adversial Intelligence and Decision-making

Post by bria » Sat Jan 30, 2016 2:34 am

I begin to answer on sensors, and then I get a little too technical (more than 1 page on only the gunshot detector...)
I drop it and began to explain the two main philosophy of getting data (swarm and central), then after writing some more, I was already at more than 2 pages....

So I will try to summarize my ideas. Don't hesitate if you want some more information on one point, I will gladly give it If I have it.
Ryan wrote:If that were indeed possible, it would change the face of CQB forever. Especially templates hostage rescue.
I don't know. RAID isn't some magic. And anything that could be used by the good guy could also , one day, be used by the bad ones.
Ryan wrote:Am I on the right track here? You need something to give you information then use the algorithm. The more information you can get, the more accurate the algorithm?
You're on the right track with some minor modification.
The more _different_ and _useful_ information you can get, the clearer the picture will be.

For example : if all your team watch the door waiting for something, You will get a very fine detailed picture of the door, but absolutely no information of the corridor right next them, where a tango is coming :)
Signal theory is exactly the same thing at work : We need multiples sensors (k9 , but also drone, indirect channel (microphone on pipe, nobody's done it ? ) ..) only because we couldn't get enough detailed information with one.

Let's take a scenario. I think it will be clearer this way.

You've got three hostage takers , they've taken twenty hostages, on two group. one of the group had a bomb with electronic captor and detonator.
They refuse all communication propositions, and close all the sheds. It's a public space, so you've got the "evacuation plan" with the blueprint of the area. Some partition were added inside the space where there are.

You have full support and can use whatever mean necessary :) (I like when it's simple).
(from now, it's only my imagination, I'm not an operator, nor I work with darpa so I could say dumb thing.)

On the support side you will deploy
- ELINT -> COMINT (without translation, but could still define their position).
- microphone, to hear what the hostage could say.
- camera under the door, or if you're sure , directly inside a partition.
- vibration microphone, on the floor.
- back-scatter x-ray (well I've said I've full support, isn't it ? .. just hope it doesn't detonate the bomb ^^)
- thermal imagery (dunno if it could see through light partition, but could help define hostage health, ...)
- ...

With all theses information you have a precise information of what is going inside, at T time.
at T+1; you've got absolutely no idea. Perhaps the three hostage takers will just surrender. Or they will start to execute hostages, or they will gather on one side of a room. You can't know.

That's where RAID is coming to help.
It will follow all theses leads, and form pattern.
A full support team could do exactly the same staff, with some limitation.
- they will tend to apply they will to the enemy. (for example they will think that "the enemy is from this cult, then it will begin to pray before killing." , but discard the fact that the enemy isn't wearing his clothes as the cult say it must. They discard an evidence because it didn't conform to the other information they had).
-> RAID doesn't have will or ego. So it could consider that the enemy decided to part from his previous cult, and automatically discard all "optimist" rules from this cult.
- Information will not be privileged on a staff-to-staff basis, following their formation to the tools.
- It could manipulate, in real time a much bigger array of information. No Information overload or over passing.
- It would not form bad habits and could detect transient change (the comint data does'nt move in more than twenty minutes, I don't need to watch it anymore. The comint change for only 0.1s after 20 min. RAID will detect it and treat it, not tag it as "noise" before even verifying).

- detect and show pretty complex and/or new pattern,
Human brain is just not fit to this sort of work

What RAID would NOT do:
- adapt and explicate directly on a mission with untrained member. It could not explicate "what the fucking red box on my screen is". Nor it could give additional precision on subject if it's not specified somewhere

From this point, all the information are centralized and process inside RAID.
Then each team could select the information they want to see. No more reason that the analyst team had the last reco picture, while the assault team is still on a 3-hour late picture, because nobody sent them the last one.

With all this, you will say "I want to assault in 20 minutes, from this and this point".
It will say in return
-> doesn't began the jamming before 14min, since they do a check at 30 min fuzzy interval fuzzy, and I calculate that they will do it with 99.9% confidence in 14min. (Nothing it's sure:) )
-> Tango 1 will be in this radius in room 1, Tango 2 will be in this radius in room 3 , Tango 3 will be in this radius in room 4", with a xxM confidence. (If they walk in a absolutely random way,this part will be with 0% confidence.But Absolute random isn't human).
-> hostage seem fine and could certainly wait for 20 min.
-> danger fire area : when you shoot Tango 1, bullet could attain group 2 hostages through a partition
-> danger : Tango 2 will be in blind spot and in position to wound hostages through partition.

Or you could use it as "best time to assault from this and this point" for example.


In this case, deploying more sensor wouldn't help because we've got the maximum precision necessary for the work (knowing where is each enemy, and how it is working with the others and the hostages.).
(in signal theory, getting more information than necessary is called "oversampling").


In addition, have you seen how nothing on the bomb come from RAID? No captor could bring information on it, so RAID will not, if something wasn't already coded inside RAID to detect it, it will not find it.
Perhaps an operator would suspect something with his instinct and warn the assault team. (but experiences shows that on MOUT, RAID was much more reliable than human instinct. Once a rule is coded, it could be applied).


Now, what to do if you get deployed on a black ops, with your team of five, and with an on foot infil of 24 hours, to capture a HVT inside a building. You're exfil is in 6h, and you've got only 4 hour to complete the assault (intelligence say you've got "only" 4 bad guys in the perimeter). You could not send a reco team inside the building, so you must adapt on the fly.

You couldn't be supported by anyone at HQ, a satellite-transfer lag is something like 2 sec.Bandwidth is scarce, and power-hungry. Then the information must be processed locally.

Multiple, light and versatile sensors array, you could drop near the target , or directly on you.
Then a local computer will aggregate and analyses all the data.
You will have a return in a clear and simple way ( I hope :) )
Of course, only basic information from the sensor and simple predictions could be showed. Mainly due to the lack of time range of the data collection, and range/precision of the sensors used).


I'm not sure that, once all hell break loose, the system wouldn't be very useful to predict anything.
It could/would still pinpoint accurately gunshot, or comint, or automatic detection of threat (for example, a moving barrel behind a barricade could be highlight by a high precision camera and showed on a HUD ) (raw data and aggregation of multiples sensors).
It could also detect pattern even when all hell break loose. For example it could detect that some call are systematically answered from another part of the building,

But all of this are purely assumption.
If I understand it well, all of theses refers to statistics on certain dimensions (capacity of the player of doing this, or that).

The idea behind RAID, I think, is a little different. It's more like the statistic to have a certain player , at a certain time, in a certain place.
It's more about how the game is played than the player constituting the game.

User avatar
Ryan
Posts: 2793
Joined: Sun Apr 10, 2011 3:10 am
Contact:

Re: Real-Time Adversial Intelligence and Decision-making

Post by Ryan » Sat Jan 30, 2016 8:14 am

Well, you're right. That was a lot of information. :lol:

I'd honestly like to see something like this in action.
CQB-TEAM Education and Motivation.

"Pragmatism over theory."
"Anyone with a weapon is just as deadly as the next person."
"Unopposed CQB is always a success, if you wanted you could moonwalk into the room holding a Pepsi."

User avatar
Ryan
Posts: 2793
Joined: Sun Apr 10, 2011 3:10 am
Contact:

Re: Real-Time Adversial Intelligence and Decision-making

Post by Ryan » Sun Apr 03, 2016 10:53 am

CQB-TEAM Education and Motivation.

"Pragmatism over theory."
"Anyone with a weapon is just as deadly as the next person."
"Unopposed CQB is always a success, if you wanted you could moonwalk into the room holding a Pepsi."

bria
Posts: 20
Joined: Sun Nov 22, 2015 11:01 pm
Location: France

Re: Real-Time Adversial Intelligence and Decision-making

Post by bria » Mon Jun 13, 2016 9:30 pm

Thanks Ryan, that a great video which is a great introduction to game theory decisions-making .

However, the showed algorithms work on a reactive way with fixed rules. You've got a fixed situation with specific rules, it will give you the odds to win or lose, then you could take a decision.

RAID seems a little more complex (well they wouldn't be darpa if it was simple ;)), since it will predict what will the next action of the ennemie(s) in real-time, on a battleground with a constantly changing rules (complex environnement, , adversaries who obey their own rules, etc...)


For a comparison, the algorithm used by google to win against go masters is fundamentally "simple" (even if state of the art) : they used multiples (hundred of thousand) go games to educated an artificial intelligence. From all these games the AI will say which is the most effective way of playing when this situation, or a similar situation, arrives.
This could be done since rules doesn't change beetween games. In addition the "teaching" and the "use" of the algorithm isn't done at the same time (no real-time teaching).
Even with all this, They use a super calculator to win a 19x19 board game with simple rules.

User avatar
Ryan
Posts: 2793
Joined: Sun Apr 10, 2011 3:10 am
Contact:

Re: Real-Time Adversial Intelligence and Decision-making

Post by Ryan » Thu Jun 16, 2016 7:32 am

Man, this is why we leave this stuff to you guys! :lol:
CQB-TEAM Education and Motivation.

"Pragmatism over theory."
"Anyone with a weapon is just as deadly as the next person."
"Unopposed CQB is always a success, if you wanted you could moonwalk into the room holding a Pepsi."

Post Reply

Return to “CQB General Discussions ”