Final Document of Ai

MILITRAY AND ARTIFICIAL INTELLEGENCE
ARTIFICIAL INTELLEGENCE:
Artificial intelligence (AI) is the intelligence of machines and the branch of computer science
NCE
INTELLENG
that aims to create it. Textbooks define the field as "the study and design of intelligent agents,"
AL
ARTIFICI
where an intelligent agent is a system that perceives its environment and takes actions that
maximize its chances of success. John McCarthy, who coined the term in 1956, defines it as "the
science and engineering of making intelligent machines”.
AI research is highly technical and specialized, deeply divided into subfields that often fail to
communicate with each other. Subfields have grown up around particular institutions, the work
of individual researchers, the solution of specific problems, longstanding differences of opinion
about how AI should be done and the application of widely differing tools. The central problems
of AI include such traits as reasoning, knowledge, planning, learning, communication, perception
and the ability to move and manipulate objects. General intelligence (or "strong AI") is still a
long-term goal of (some) research.
MILITARY:
A military is an organization authorized by its nation to use force, usually including use of
weapons, in defending its country (or by attacking other countries) by combating actual or
perceived threats. As an adjective the term "military" is also used to refer to any property or
aspect of a military. Militaries often function as societies within societies, by having their own
military communities, economies, education, medicine and other aspects of a functioning civilian
society.
In the modern era, world wars and countless other major conflicts have changed the employment
of the militaries beyond recognition to their ancient participants. Empires have come and gone;
states have grown and declined. Enormous social changes have been wrought, and military
power continues to dominate international relations. The role of the military today is as central to
global societies as it ever was.
History:
Ever since the invention of the catapult, warfare has been the main driving force behind the
advancement of technology. The military has played a main role in researching new technologies
NCE
INTELLENG
and their work has led the way for significant advancements in other fields. Many great
AL
inventions, including the Internet, have been made from military funded research. The driving
ARTIFICI
force has been the escalation of weaponry. If you don't have something your enemy does, you are
very often at a disadvantage. Computers were first developed to calculate missile trajectories and
break enemy codes. With this in mind it can be said that the military has been and will continue
to be AI's main driving force. The use of such technology opens up endless possibilities in the
military. Currently AI robots are mainly used to find roadside bombs, search caves, and act as
armed sentries, and they have the potential to do so much more.
Definition:
AI in the military is defined as any system which uses the help of ‘intelligent' machines to
complete tasks. When people think of Artificial Intelligence in this area, images of Terminator-
like killing machines automatically spring to mind. However, not all AI relates directly to the
battlefield and it can server much more passive purposes.
For example, the US Military has developed a computer game that uses AI to teach new recruits
how to speak Arabic. In the program soldiers complete game missions where they must be able
to understand and speak the language to succeed. The game works by using speech recognition
technology to evaluate the soldier's words and identify common errors.
Another definition:
The obvious question now is: just how real is artificial intelligence? The hype has been high for
years about what is to come, what is almost here, and what tremendous things AI will do, once
we actually have it. But is AI real or is it just another form of vaporware?
The answer, according to researchers at Pathfinder Systems Inc. (Lakewood, Colo.), is ... mixed.
"An extreme case of these hopes, in our sphere of interest, is recognition of the evolving battle
scene and directly generating doctrinally correct plans, operations orders, or frag orders from this
scene recognition," according to a Pathfinder analysis of the current state and future prospects of
AI. "While human 'command agents' indeed appear to work this way, we believe there is very
NCE
INTELLENG
little chance of achieving this in the near future, either by neural nets or by 'classical AI'
AL
ARTIFICI
methods.
"The fact is that the processing and storage power available in today's computers is wholly
inadequate to approximate a 'neural net' model of the human brain's functions," the Pathfinder
analysis goes on. "The human brain uses around 10 billion times less power per 'instruction' than
modern processors; it has perhaps 100 to 300 trillion bytes of equivalent 'main RAM memory'
and it learns about 1 million to 1 billion times faster than our best neural net or other software
models." In contrast, a human child seems to generate approximately 1 million new neural
connections per second — or perhaps 1 megabyte of new code per second.
Pathfinder forecasts the hardware necessary for human-like AI, if the current explosive growth in
computing technology continues, evolving as follows:
• if only memory were the limiting factor, we would have AI around 2050;
• if heat/instructions were the limiting factor, we would have AI around 2070; or
• With learning were the limiting factor, we could not have AI before 2160.
Of course, researchers can do much in the realm of sorting, identifying, processing, and fusing
crucial data — tasks that fall far short of human-level intelligence. It may be 2160 before AI
reaches a level where it alone can make battlefield decisions — a prospect that human
commanders probably will never welcome. But the ability to provide accurate, crucial data in
real-time in a form those same human commanders can use to help them make battlefield
decisions themselves should be much closer to reality.
In a paper on "Intelligent Mobile Agents in the Military Domain", scientists at Lockheed Martin
agree that much remains to be done, including meeting a variety of technical requirements to
prove and support widespread transition of agent technology into the military arena.
"Robust agent behavior control mechanisms must be developed and emplaced; the military must
be convinced that agents are controllable tools rather than dangerous (in terms of security and
bandwidth) and uncontrollable viruses. Conversely, resource control mechanisms must be
instituted because agents have greater potential than human operators to overload legacy
NCE
INTELLENG
resources," according to the Lockheed Martin paper. AL

ARTIFICI
"For their part, military network designers must carefully define data and legacy-system access
and release policies so that agents are not in danger of inadvertently disseminating sensitive
information. From an agent technology standpoint, perhaps the most outstanding need is for a
standard agent capability and data description language — a semantic framework that supports
collaboration across many heterogeneous agent systems. DARPA's Control of Agent Based
Systems program has put in place a JINI-based software infrastructure to support agent control
and collaboration."
JINI is not an acronym, but the name of a Java network technology of Sun Micro Systems.
Kolodzy's program at DARPA is to devise an integrated mobile communication system with fast
data throughput, optimized for restrictive terrain and using several different waveforms with self-
configuring networking. According to DARPA, the top technical challenges they must meet
before having a system suitable for field use by the Army, Marines, and Special Operations
Forces are:
• low probability of intercept/low probability of detection (LPI/LPD) waveforms;

• |mobile ad hoc networking; and
• Position and navigation in GPS-denied areas.
"SUOSAS was developed because individual war fighters are going to be disbursed in the future,
as noted by a Defense Science Board report back in 1996," Kolodzy says. "So a lot of situational
awareness information must be sent out to the individual war fighter. Some people refer to it as
'bits over iron’; meaning bits sometimes have more power than iron. But it is hard to get a radio
onto an individual war fighter and make sure no one else can hear him."
Military Organization & Training:

Most early Greek city-states did not having standing, or full time, armies. Instead, their armies
consisted of citizens who were required to serve only in times of war. All soldiers had to provide
their own weapons and armor, as well as enough food for several days in the field.
NCE
INTELLENG
Because they were not professional soldiers, the Greek soldiers did not have as advanced skills
AL
as a professional soldier would. In Athens, schools known as gymnasia were designed to develop
ARTIFICI
the physiques of your boys in preparation for their skills as citizen soldiers. All boys began their
military training at the age of 18. The training eventually included a two year tour duty at a
military post. All Athenian males up to the age of 60 could be called up for active military
service. Spartan boys were taken from their families at the age of 7 to begin military training.
They passed through various military organizations until they reached adulthood and entered
military units officially
Artificial intelligence (AI) is getting down to business. The field is moving away from the
grandiose attempts of the past that attempted to reproduce the spectrum of human intelligence in
machine form. Now the focus is on specific applications employing AI principles to tackle
relatively mundane tasks. This is a natural evolution based on the availability of rapidly
accelerating raw computer power. In fact, AI techniques are becoming so ubiquitous that the
computers that now bear the label "Intel inside" could well be labeled "AI inside," says Alan
Meyrowitz, director of the Navy Center for Applied Research in Artificial Intelligence at the
Naval Research Laboratory (NRL) in Washington.
Navy experts say AI addresses the automation and extension of human intellectual skills —
specifically human decision making and autonomous devices. Both have their civil and military
counterparts.
In the former category, decision making, the military employs AI for planning and logistics
functions. In the commercial world similar techniques help business leaders streamline retailing
and financial operations by extracting relevant data — known in the AI business as data mining.
Autonomous devices, such as the robotic systems in automobile assembly lines, were well
established in manufacturing industries before U.S. military experts launched their theoretical
studies under the Strategic Computing Program at the Defense Advanced Research Projects
Agency (DARPA) in the 1980s. Now, with a combination of military-funded development

programs and the availability of commercial off-the-shelf (COTS) technology, the military
services are beginning to implement AI methods in such new generations of weapons platforms
as unmanned aerial vehicles (UAVs) and autonomous submersibles to perform unmanned
NCE
INTELLENG
surveillance in shallow waters. Robots for military and civil applications, such as the grippers on
AL
ARTIFICI
factory assembly lines, are even beginning to look anthropomorphic, Meyrowitz says.
This means that "AI has moved from science in the laboratory to technology built for useful
things," Meyrowitz says. Moreover, two of the principal thrusts — data mining and machine
vision — have taken on identities of their own outside the AI umbrella.
So-called "expert systems" are part of a rich AI history. In their simplest forms, this computing
technique uses a data base of rules and rudimentary "if-then" logic processing to help humans
determine the best courses of action under certain sets of circumstances.
Where do expert systems fit into the AI picture today? Much touted in the past in such programs
as DARPA's Pilot's Associate, modern expert systems are evolving in two directions: what
Meyrowitz calls rule-based and case-based. "All expert systems are AI, but not all AI is expert
systems," he comments.
The traditional rule-based systems, which were supposed to capture the best expertise of pilots
and other skilled personnel, tend to be more detailed and smaller in their scope of information,
Meyrowitz says.
Case-based reasoning, which is the way humans learn, involves storing previous problems and
solutions and then identifying a match when a new situation arises. Meyrowitz calls this
approach "more abstract and higher level," as case-based expert systems extract rules from cases
over time. Some computer scientists argue that this capability represents the beginning of
"machine learning." Critical to the ability to employ this method is the availability of large,
affordable computer memory — and this is where AI is riding the crest of today's commercial
electronics technology.
This capability, in turn, raises the perennial question of man-in-the-loop. NASA leaders have
wrestled with this issue since the inception of the space agency in 1958 in trying to determine
which space missions should be manned and which unmanned.
NCE
INTELLENG
Hans Moravec, the robotics expert at Carnegie Mellon University in Pittsburgh, has been an
AL
outspoken advocate of using as much machine automation in space as possible — including on
ARTIFICI
the International Space Station. Moravec supports machine automation on the grounds that
machines are cheaper to support in space than are humans since they do not require expensive
life support. Conversely, humans still outperform robots on Earth except in applications that are
time-critical or hazardous, he says.
Military service leaders face a comparable problem of sorting out how much robotics they want.
Meyrowitz says the long-term challenge is to evolve toward automation in battle "with
confidence."
Here is where COTS technology can help. Meyrowitz divides the issue into "advisory" and
"automatic" applications of AI, and cites examples from the commercial world that could be
valuable to the military.
The medical profession uses advisory AI techniques to match patients' symptoms with a set of
possible actions to improve treatment. Automatic AI helps control industrial processes, such as
shutting down an overheating furnace.
The reality is that man is always in the loop in what Meyrowitz calls a "continuous spectrum of
various degrees of connection." He likens this to the division of labor in an office. And, like an
office or any other workplace, systems must link into a network so they can work together. The
consequent distributed AI systems exhibit what he calls a variation of social behavior.
The problem with AI in the past is that it was oversold, notes Dave McDaniel, president of Silver
Bullet Solutions in Arlington, Va. "In the 1980s expert systems were all the rage," he says, but
today the focus is on specific — if less glamorous — tasks.
McDaniel stresses the need for AI techniques in sensor fusion. Driven by dramatic improvements
in data storage and communications links, these techniques are becoming available to the
operators of fielded military systems.
NCE
INTELLENG
For example, he points out, the Airborne Warning and Control Systems (AWACS) aircraft,
AL
Aegis warships, and reconnaissance satellites are all picking up signals from targets. The
ARTIFICI
problem is with the ambiguity of those data.
Expert systems can help the operators by estimating the locations and intentions of enemy forces.
Since such AI-based systems are likely to be networked together, this means that all the forces in
a theater of operations can have a common operational picture. McDaniel cites the Navy's
Cooperative Engagement Capability — better known as CEC — as an example.
Moreover, the same AI techniques employing expert systems and natural languages that drive the
robotic systems in UAVs and autonomous ground vehicles share a technology base with
industrial factories, he adds. In each case, the need for interpreting data and taking the
appropriate action is functionally similar.
The COTS angle is born out by the experience of Steve Chien, principal scientist for automated
planning and scheduling in the artificial intelligence group at the Jet Propulsion Laboratory (JPL)
in Pasadena, Calif. "There have been a lot of AI success stories in recent years," he says, "but
you hear less about the government applications because they are dwarfed by the commercial
market.
"AI has taken off since 1995," Chien continues. He calls the situation an "explosion." He focuses
on machine learning, knowledge discovery (which he calls a community of its own), and data
mining. "Every company needs to do data mining, but how many science applications are there?"
JPL experts are implementing what is known as the Automated Scheduling and Planning
Environment system, or ASPEN, as a sort of command and control architecture in which
participating scientists can do data mining to achieve sensor fusion. It in effect gets all the
participants on the same page so they can tell NASA what data products they want. The
spacecraft knows the constraint parameters and automatically generates the data according to
plan.
Experts credit ASPEN with reducing the acquisition planning time from a three-year process to
NCE
INTELLENG
one week for an Antarctic mapping mission conducted in September. Furthermore, ASPEN is to
AL
be used in an Explorer spacecraft mission conducted by the University of Colorado and
ARTIFICI
scheduled for launch later this month.
The ultimate goal is the autonomous spacecraft, but Chien concedes, "We're not quite there yet."
On the ground, however, JPL has used AI techniques effectively to improve the efficiency of its
network.
One example is what JPL officials call their operations mission planner for its 26-meter antenna
subnetwork. They have used AI since 1993 and have reduced the scheduling effort by 30 percent
while doubling network support, according to Chien.
Another is an automated image processor for synthetic aperture radar data, which has cut manual
inputs by a factor of 10 and reduced the amount of central processor time necessary for
processing by 30 percent.
Also at JPL a program known as Continuous Activity Scheduling, Planning, Execution and
Replanning, or CASPER, is the real-time version of ASPEN that does planning and execution,
and is designed for flight use as part of the flight software. CASPER is for another University of
Colorado mission scheduled for launch in the summer of 2002.
These planning and scheduling systems share something in common with the commercial world
— they improve the logistics management, Chien says. "How do you get your groceries to the
supermarket? Where do you produce? How do you ship?" he asks. Those are the kind of
questions JPL managers ask, too, and Chien says COTS hardware is so readily available as to
make hardware availability an irrelevant concern.
There are also COTS opportunities in knowledge discovery and data mining, he adds. The
commonality here — and hence the available technology base — derives from such widespread
commercial practices as the bar-coded merchandise in supermarkets. A principal objective in this

application is to determine buying patterns to reduce costly inventory.
Similar methods are part of financial markets to help experts understand trends and direct-mail
NCE
INTELLENG
marketing to identify customers and thus customize e-mail marketing lists. AL
ARTIFICI
Looking to the future, Chien speculates that AI can help achieve automatic target recognition, or
ATR. Yet the big need is to solve the sensor fusion problem. There is overlap here between the
military's needs for UAVs and autonomous ground vehicles planned for urban warfare situations
and the space scientists' needs for autonomous vehicles such as the recent Mars rover. "The
spacecraft is the ultimate UAV," he says, and AI will play a key role in training.
Further evidence of the dual-use aspects of AI technology comes from Hatte Blejer, vice
president and director of intelligent information systems at SRA International in Fairfax, Va. Her
company is principally involved in providing intelligent information systems for civilian
government agencies, such as the National Institutes of Health. She says she sees a logical
progression of data mining into military applications.
She says the focus now is on recognizing sequences of events and discovering patterns. Digital
text tends to be unstructured, and computers traditionally have needed structured information. In
the past five years, according to Blejer, data mining has done just that.
A system that SRA experts developed for NASDAQ four years ago detected new patterns of
fraud by perpetrators who, like enemy missiles, "fly below the radar." Without data mapping
these incursions would go undetected, Blejer maintains. "There are not enough human eyes." The
company has also worked on physical intrusion-detection systems for the military.
Moreover, these new data-mapping systems are hardware- and software-independent, she says:
"NT, Unix, and Linux, whatever." All the hardware and software are readily available off the
shelf. The trick is to integrate all the components and thus add value in business-to-business as
well as military applications.
Another facet of AI — artificial neural networks — is also susceptible to dual use, according to
Gary Layton, marketing vice president at interBiz, a division of Computer Associates based in
Islandia, N.Y. He says a neural network is software that functions much like the human brain in
its capacity to learn, accumulate knowledge, and apply this knowledge to new situations.
NCE
INTELLENG
AL
Yet this requires substantial processor and memory assets, and the many of the new sophisticated
ARTIFICI
neural network applications would not be possible without the current advances in the electronics
industry. The company's package, known as Neugents, is based on this computer power.
Layton cites such diverse applications for the package as enabling chemical companies to
improve their methods of mixing chemicals and the U.S. Army to analyze the causes of engine
failure in tanks and thus improve the mean time between failures.
Navy evaluates shipboard network to coordinate
OWENS MILLS, Md. — A form of artificial intelligence is coordinating scheduling information

aboard the guided missile destroyer USS McFaul (DDG 74) to coordinate scheduling
information on the Norfolk, Va.-based Arleigh Burke-class warship.
A data synchronization software package known as ScoutWare from Aether Systems Inc. of
Owens Mills, Md., ties together Palm hand-held terminals from Palm Inc. in Santa Clara, Calif.,
with a server from Clarinet Systems in San Jose, Calif. The ship's crewmembers use the system
to send and receive e-mail, conduct training and evaluations, consolidate checklists and
databases, and coordinate schedules.
The wireless system is initially limited to about 150 officers and sailors using the Palm devices
on the destroyer, and the purpose is to improve response times to boost combat readiness.
By coordinating sensor data into a shared network, the crewmembers should have a better picture
of the operations under way. The McFaul is one of the ships able to launch Tomahawk cruise
missiles, and it also performs anti-submarine warfare missions.
All the hardware is commercial off-the-shelf, and the Navy is using the McFaul as the test ship
for possible application of the software package throughout the Atlantic Fleet.
The U.S. Army has also been considering this technology for use with its medical research
NCE
INTELLENG
databases, and the Air Force is looking into it for inspections of aircraft on the flight line. — J.R.
AL
ARTIFICI
AI techniques at Air Force Research Lab to extend target identification
DAYTON, Ohio — one of the ways artificial intelligence can extend military capabilities is in
interpreting "non-literal sensors," says Dale Nelson, chief of the target recognition branch in the
Sensors Directorate of the Air force Research Laboratory at Wright-Patterson Air Force Base in
Dayton, Ohio.
"People don't see in X-ray or listen to sonar signals," he says, and that is a role he envisions for
AI in target recognition. Synthetic aperture radar, or SAR, is not like anything on the human
body. "SAR looks like a photo, but it's not," Nelson notes.
Yet the medical community uses similar AI techniques for X-ray interpretation, and the military
can put this technology to work for its unique missions, such as distinguishing a tank from a
school bus.
Nelson, who quips that AI is "what you haven't done yet," is investigating control programs
based on what are known as genetic algorithms to enable aircraft to get smarter after each battle.
Like the human genes for which they are named, genetic algorithms permit knowledge to be
passed down from one generation of weapon system to another.
The effort also involves data mining. "The Air Force, like companies, has vast databases and we
need to get useful patterns," Nelson says. The approach is to organize the data into a table in
which each column is a different attribute of the target and each row is the target.
By using AI techniques to reduce the number of columns, the idea is to find the minimum
number of features to identify all targets. He estimates this data reduction can cut the 128
candidate attributes to about 25 important attributes for target recognition.
"We've picked the low hanging fruit, and now we need automatic learning," Nelson notes. "The
organism that doesn't learn is dead." — J.R.
Army scientist: computer power not the answer to all AI needs NCE
INTELLENG
AL
FORT MONMOUTH, N.J. — Raw computing power alone is not the answer to applying
ARTIFICI
artificial intelligence techniques to the military's needs, says a U.S. Army software expert.
Instead, specialists need a combination of AI techniques, such as neural nets and expert systems,
says Gerald Powell, deputy director for the operations directorate at the Software Engineering
Center of the Army's Communications and Electronics Command (CECOM) at Fort Monmouth,
N.J.
The need for a combination of computational technologies is especially true for the Army's
logistics planning to support the ammunition and petroleum needs of deployed forces, Powell
says. Moreover, he insists, this technology can be shared with the commercial sector.
Despite the big AI push among the services and the Defense Advanced Research Projects
Agency in the 1970s and '80s, Powell discerns a slowdown in the rate of research these days as
the commercial sector takes the lead and the military attempts to focus on its unique
requirements.
The solution he envisions is integrating such technologies as neural networks and genetic
algorithms to tackle pressing problems like automatic target recognition. "We need to know what
types of tanks, enemies and friendlies, are on the battlefield to avoid fratricide," Powell says.
As an example of why raw computing power is not the whole solution, he cites a 1956 study by
Dartmouth University that found in the game of chess there were 10 to the 120th power possible
moves. This is far beyond the capability of any human — or even any conceivable
supercomputer that could be developed.
To illustrate that level of complexity, physicists estimate that the total number of subatomic
particles in the universe is about 10 to the 85th power. Or, to put it even more dramatically, if
there were 10 to the 86th power of elementary particles, there would be nine more universes just
like this one. — J.R.
NASA robot in form of snake planned to penetrate inaccessible areas NCE

INTELLENG
AL
MOUNTAIN VIEW, Calif. — NASA engineers are developing a new type of snake-shaped
ARTIFICI
robot at the NASA Ames Research Center in Mountain View, Calif., to explore areas where a
wheeled robotic rover might get stuck or topple over.
Engineers have built a mechanical prototype of what NASA leaders are calling the "snakebot,"
says Gary Haith, lead engineer on the project. Yet the NASA team is now working on the
sensors and microcontrollers for operational use. Team members also plan to write software to
enable the device to learn by experience in crawling over various surfaces and climbing over
obstacles.
The NASA center is working with the nearby Xerox Palo Alto Research Center in Palo Alto,
Calif., where Mark Yim developed a slightly different version known as the "polybot." The next
step is to simulate the snakebot in a computer program to develop control procedures.
The purpose is not to replace NASA's existing wheeled robots, such as those used to explore the
surface of Mars, but to complement them with a smaller, cheaper device that can operate
independently in tight places. One of the advantages is the robot's ability to crawl off a spacecraft
lander without a ramp, Haith says.
Military Applications of AI:
The military and the science of computers has always been incredibly closely tied - in fact, the
early development of computing was virtually exclusively limited to military purposes. The very
first operational use of a computer was the gun director used in the second world war to aid
ground gunners to predict the path of a plane given its radar data. Famous names in AI, such as
Alan Turing, were scientists that were heavily involved in the military. Turing, recognized as one
of founders of both contempory computer science and artificial intelligence, was the scientist
who broke the German's Enigma code through the use of computers.
As computing power increased and pragmatic programming languages were developed, more
NCE
INTELLENG
complicated algorithms and simulations could be realized. For instance, computers were soon
AL
utilized to simulate nuclear escalations and wars or how arms races would be affected by various
ARTIFICI
parameters. The simulations grew powerful enough that the results of many of these 'wargames'
became classified material, and the 'holes' that were exposed were integrated into national
policies.
Artificial Intelligence applications in the West started to become extensively researched when
the Japanese announced in 1981 that they were going to build a 5th Generation computer,
capable of logic deduction and other such capabilities.
Inevitably, the 5th Generation project failed, due to the inherent problems that AI is faced with.
Nevertheless, research still continued around the globe to integrate more 'intelligent' computer
systems into the battlefield. Emphatic generals foresaw battle by hordes of entirely autonomous
buggies and aerial vehicles, robots that would have multiple goals and whose mission may last
for months, driving deep into enemy territory. The problems in developing such systems are
obvious - the lack of functional machine vision systems has lead to problems with object
avoidance, friend/foe recognition, target acquisition and much more. Problems also occur trying
to get the robot to adapt to its surroundings, the terrain, and other environmental aspects.
Nowadays, developers seem to be concentrating on smaller goals, such as voice recogition

systems, expert systems and advisory systems. The main military value of such projects is to
reduce the workload on a pilot. Modern pilots work in incredibly complex electronic
environments - receiving information not only from their own radar, but from many others
(principle behind J-STARS). Not only is the information load high, the multi-role aircraft of the
21st century have highly complex avonics, navigation, communications and weapon systems. All
this must be organized in a highly accessible way. Through voice-recognition, systems could be
checked, modified and altered without the pilot looking down into the cockpit. Expert/advisory
systems could predict what the pilot would want in a given scenario and decrease the complexity
of a given task automatically.
Aside from research in this area, various paradigms in AI have been successfully applied in the
NCE
INTELLENG
military field. For example, using an EA (evolutionary algorithm) to evolve algorithms to detect
AL
targets given radar/FLIR data, or neural networks differentiating between mines and rocks given
ARTIFICI
sonar data in a submarine. I will look into these two examples in depth below.
Genetic Programming:
Genetic programming is an excellent way of evolving algorithms that will map data to a given
result when no set formula is known. Mathmaticians/programmers could normally find
algorithms to deal with a problem with 5 or so variables, but when the problem increases to 10,
20, 50 variables the problem becomes close to impossible to solve. Briefly, how an GP-powered
program works is that a series of randomly generated expression trees are generated that
represent various formulas. These trees are then tested against the data, poor ones discarded,
good ones kept and breed. Mutation, crossover, and all of the elements in genetic algorithms are
used to breed the 'highest-fitness' tree for the given problem. At best, this will perfectly match
the variables to the answer, other times it will generate an answer very close to the wanted
answer. (For a more in-depth look at GP, read the case study)
A notable example of such a program is SDI's e evolutionary algorithm designed by Steve Smith.
e has been used by SDI to research algorithms to use in radars in modern helicopters such as the
AH-64D Longbow Apache and RAH-66 Comanche. e is presented with a mass of numbers
generated by a radar and perhaps a low-resolution television camera, or FLIR (Forward-looking
Infra-red) device. The program then attempts to find (through various evolutionary means) an
algorithm to determine the type of vehicle, or to differentiate between a actual target and mere
"noisy" data.
Basically, the EA is fed with a list of 42 different variables collected from the two sensors, and
then a truth value specifying whether the test data was clutter or a target. The EA then generates
a series of expression trees (much more complicated than those normally used in GP programs).
When new a best program is discovered, the EA uses a hill-climbing technique to get the best
possible result out of the new tree. Then, the tree is subjected to a heuristic search to optimize the
tree.
Once the best possible tree is found, e will output the program as either pseudocode, C, Fortran
NCE
INTELLENG
or Basic. AL
ARTIFICI
Once the EA had evolved the training data, it was put to work on some test data. The results
were quite impressive:
Percent Errors
Sensor Training Data Test Data
Radar 2.5% 8.3%
Imaging 2.0% 8.0%
Fused 0.0% 4.2%
While the algorithms performed well on the training data, the performance increased a lot when
applied to the test data. Nevertheless, the fused detection algorithm (using both radar and FLIR
information) still provided a decent error percentage.
An additional plus to this technique is that the EA could be actually programmed into the
weapon systems (not just the algorithm outputted), so that the system could dynamically adapt to
the terrain, and other mission-specific parameters.
Neural-networks:
Neural networks (NN) are another excellent technique of mapping numbers to results. Unlike the
EA, though, they will only output certain results. A NN is normally pre-trained with a set of
input vectors and a 'teacher' to tell them what the output should be for the given input. A NN can
then adapt to a series of patterns. Thus, when feed with information after being trained, the NN
will output the result whose trained input most closely resembles the input being tested.
This was the method that some scientists took to identify sonar sounds. Their goal was to train a
network to differentiate between rocks and mines - a notoriously difficult task for human sonar
operators to accomplish.
NCE
INTELLENG
The network architecture was quite simple, it had 60 inputs, one hidden layer with 1-24 inputs,
AL
and two output units. The output would be <0,1> for a rock and <1,0> for a mine. The large
ARTIFICI
amount of input units was to encorporate 60 normalized energy levels of frequency bands in the
sonar echo. What this means is that a sonar echo would be detected, and subsequently fed into a
frequency analyzer, that would break down the echo into 60 frequency bands. The various
energy levels of these bands was measured, and converted into a number between 0 and 1.
A few simple training method was used (gradient-descent), as the network was fed examples of
mine echoes and rock echoes. After the network had made its classifications, it was then told
whether it was correct or not. Soon, the network could differentiate as good or better than its
equivalent human operator.
The network had also beaten standard data classification techniques. Data classification
programs could successfully detect mines 50% of the time by using parameters such as the
frequency bandwidth, onset time, and rate of decay of the signals. Unfortunately, the remaining
50% of sonar echoes do not always follow the rather strict heuristics that the data classification
used. The networks power came in its ability to focus on the more subtle traits of the signal, and
use them to differentiate.
Morality: A Quick Thought:

All these systems are quite impressive, and perfected models could prove incredible assets on the
battlefield. Artificial Intelligence may only get developed to a certain level due to the threat
humans feel as computers get more and more intelligent. The concepts behind movies such as
Terminator where our robotic military technology backfires on us and destroys us are rampant.
Are there moral issues that we must confront as artificial military intelligence develops? As Gary
Chapman puts it:
Autonomous weapons are a revolution in warfare in that they will be the first machines given the
responsibility for killing human beings without human direction or supervision. To make this
more accurate, these weapons will be the first killing machines that are actually predatory, that
are designed to hunt human beings and destroy them.
The applications of AI in the military are wide and varied, yet due to the robustness, reliability,
NCE
INTELLENG
and durability required for most military programs and hardware, AI is not yet an intregral part
AL
of the battlefield. As techniques are refined and improved, more and more AI applications will
ARTIFICI
filter into the war scene - after all, silicon is cheaper than a human life.
Usage of AI in military:
 Example:
The Generalized Weapon Target Assignment Problem
1. Introduction:
A key component in planning and dynamic control of missions is the assignment of resources
(e.g., different aircraft types and weapons) to targets. The Weapon Target Assignment (WTA)
problem is to find a proper assignment of platforms/weapons to targets with the objective of
maximizing the overall effect associated with targets. Various methods for solving this NP-
complete WTA problem have been reported in the literature [1-6]. These studies have focused
weapon pairing aspects or sortie analysis with regard to targets. In this study, we consider one or
more types of weapons carried by a set of platforms against a set of targets, and extend the basic
WTA problem by allowing for multiple target assignments per platform. We also investigate
how the formulation can be applied to collaborative planning where multiple sources may be
required per target.
2. Model
This section describes the integer programming formulation of the generalized weapon target
assignment problem. Suppose a ground command center or an airborne mission command center
has to reassign a set of platforms or reallocate their weapons to a set of targets. Each platform is
assumed to carry one or more weapon types, and each target is fully or partially satisfied by one
type of weapon served by a platform (referred to as the “source” hereafter). A source is able to
serve multiple targets, and we assume it delivers only one type of weapon when it visits a target.
A source leaves the starting location, serves different targets and returns to the ending location.
The source cannot travel over the distance of its travel capacity (based on available and bingo
fuel). An assignment (or task) can be described as assigning a source to targets with proper
weapons satisfying the travel capacity limit. We assume that weapon effectiveness factors (range
NCE
INTELLENG
from 0 to 1) related to the specified quantities of a weapon type and target type are given.
AL
ARTIFICI
Targets are also assigned values to reflect their significance and priority. The benefit of an
assignment, which considers each source-target-weapon combination, may be written as
Benefit = Value x (Weapon Effectiveness) x Plength where Plength = M / (M + distance), and M
is a constant
For example, as shown in Figure 1, source 1 starts at position (1,1), serves target 1 at position
(3,7) and target 2 at position (6,8), and ends at (2,1).
Figure1. A sample task for source 1
Suppose source 1 has 20 travel capacity units and four weapons of type A, two weapons of type
B, and two weapons of type C. Suppose target 1 has a target value of 200 and needs to be served
by either one weapon of type A with a combined effectiveness 0.8, one weapon of type B with an
effectiveness of 1.0, or two weapons of type C with a combined effectiveness of 0.9. Similarly
suppose target 2 has a target value of 100 and needs either two weapons of type A with
combined effectiveness 1.0, or one weapon of type B with effectiveness 0.9. One of possible
assignments for source 1 can be starting from (1,1), serving two type C weapons for target 1 and
two type A weapons for target 2, and returning (2,1). The benefit from serving target 1 is equal
to 179.04 and the benefit from serving target 2 99.68, where M = 1000. Thus the assignment
yields total benefit 278.72 with traveling 16.61 distance units. The objective of the Generalized
Weapon Target Assignment Problem is to maximize the total benefit by selecting the best set of
assignments for the sources. Suppose that the assignment problem has m sources and n targets.
NCE
INTELLENG
Then the problem may be formulated as AL

ARTIFICI
where, J is the set of all feasible assignments and cj the total benefit from assignment j, j ∈ J ; s
S , s = 1,…,m, is the subset of assignments to which source s is assigned; t T , t = 1,…,n, is the
subset of assignments that serve target t. This formulation is similar to the vehicle routing
problem.
3. Enumeration of Assignments
This section describes how to enumerate the set of all possible assignments J. From target
information we can enumerate all target-weapon-quantity sets. For instance, target 1 in the
example from Section 2 needs to be served with either one weapon of type A, one weapon of
type B, or two weapons of type C. Each one of these weapon and quantity combinations becomes
a demand waiting for a proper source, which is called a target drop. Therefore, target 1 has three
target drops, i.e., target 1-A-1, target 1-B-1 and target 1-C-2. In this manner we can enumerate
all possible target drops from target information. From source and target information we can set
all possible assignments, and each of them is composed of a source and sequence of target drops,
called a target drop set. Each of target drop set can then be combined with a source, called an
assignment. We can set all possible assignments from source information and target drop sets,
and each assignment yields benefit from assigning a source to a target drop set as described in
Section 2. Out of these possible assignments we can select feasible assignments, which satisfy
the travel capacity limit of related sources.
4. Branch and Bound Algorithm

NCE
INTELLENG
There are many ways to solve an integer programming problem, and we adapted a simple branch
AL
ARTIFICI
and bound method to handle the weapon target assignment problem because of its flexibility.
This section describes the branch and bound algorithm, which is similar to implicit enumeration.
Instead of enumerating all possible assignments the algorithm deletes unnecessary enumeration
steps and improves its efficiency. Greedy method, skipping unnecessary branching and lower
bound rules are used, which will be explained later in this section. We can see how the branch
and bound algorithm works by using enumeration tree in Figure 2. Each node represents a
variable, and each branch is a value of a variable. The tree represents four variables and all
possible branches. The bottom nodes show all set of solutions. In Figure 2 node 1 implies that
X1=1, X2=1, X3=1, and X4=1, and node 2 similarly indicates X1=1, X2=1, X3=1, and X4=0
and so forth. If formulation has four assignment variables, then there are sixteen possible
solutions. However, some of them are not feasible because of constraints (2) and (3). Moreover,
some solutions are not as good as the others. These properties enable the branch and bound
algorithm to delete many of nodes and of branches as shown in Figure 3.
NCE
INTELLENG
AL
ARTIFICI
We arbitrarily choose to branch to the left first at each node in Figure 3. Suppose that node 1 is
infeasible because of a source conflict constraint (2), so we branch to the right. We can keep
branching until we reach node 2 and find a feasible solution. This solution provides a lower
bound for other solutions. Because the objective function is nonnegative the solution at node 3 is
inferior to the one at node 2. Suppose that node 4 is infeasible because of a constraint; then we
NCE
INTELLENG
need to backtrack until we reach the root node (top node) to branch further. The algorithm might
AL
enumerate all possible assignments. In this case, the total number of nodes is 2N+1 –1, so it
ARTIFICI
would not be practical for large N. The following methods help the algorithm to reduce the
number of nodes created by branching recursively and to find (near) optimal solutions as soon as
possible. After enumerating all feasible assignments described in Section 3, we sort these
assignments in decreasing order of their benefits.
The reason for doing this is that we might find a near optimal solution quickly. Suppose the
assignments are sorted, and nodes for X1, X2, X3, and X4, in Figure 2, represent the assignment
with maximum benefit for the first, second, third, and fourth source, respectively. If node 1 is
feasible, then it is optimal because the assignment’s benefit is greater than all of assignments for
those sources. However, the left-most branch is rarely feasible, so we can select feasible
assignments greedily until a feasible solution is found. Because of the greedy selection, the
solution is likely near optimal, so we will be able to prune several suboptimal solutions and limit
the size of the branch and bound tree.
5. Sequential Method
One heuristic method to solve the generalized weapon target assignment problem is to assign
sources to targets sequentially. Given n sources and m targets, we set up a directed bipartite
graph (G) of sources and targets, and look for the solution of the
Asymmetric assignment problem:

NCE
INTELLENG
AL
ARTIFICI
In the above, xij represents a single assignment of source i to target i, using the resource with
the highest effectiveness for that target, and aij the corresponding benefit for this pairing. (In the
example from Section 2, this would be Target Drop 1-B-1, when source
1 is to be assigned to target 1). We use the auction algorithm [7] to solve the above problem.
Once the primary assignments are identified, we look for secondary assignments for sources with
unallocated resources against any unassigned targets, and keep repeating the bipartite graph
build-and-auction process until no feasible assignments remain. This greedy approach is much
faster compared to the branch-and-bound scheme, as we enumerate and investigate but only a
very small subset of the assignments for the latter (roughly O(nm) vs. O(nm2) ).
7. Extension of branch-and-bound method to multiple source assignments per target In section 4

we discussed the case where a target is assigned to a single source. Consider the following
trivial example, where the assignments are sorted by source, in decreasing benefit order:
NCE
INTELLENG
AL
ARTIFICI
When we allow more than one source per target, e.g. two platforms each delivering half the total
required weapons for a desired effectiveness, an assignment will have pointers two multiple
sources. Modifying the example above, where we consider assignment X5 to contain source 2, in
addition to source 3, we have
NCE
INTELLENG
AL
ARTIFICI
where the branches from X2 to X5, X3 to X5, X4 to X5, and X5 to X6 are automatically
eliminated as infeasible branching. The branch X5 to X5 from source 2 to source three is a fixed
branch, dictated by the source constraints.
7. Computational Experiments
We compared the solutions found using the branch-and-bound method with those from the
sequential method using random instances from a Monte Carlo Simulation. We then considered
the following parameter settings:
NCE
INTELLENG
AL
ARTIFICI
Max time is the maximum amount of time we allowed the branch-and-bound code to execute,
and objective tolerance is used to prune nodes in the branch-and-bound tree that are within a
tolerance of the best known feasible solution. By modifying one parameter, we generated
fourteen different parameter sets. In each parameter set was used to solve 50 instances randomly
generated instances. The average benefit of the corresponding 50 solutions by the branch-and-
bound method is compared with that of the sequential method.
From Table 4, the best improvement from the sequential method to the branch-and- bound
method is in assigning multiple sources to targets. In this case, the average benefit from using the
NCE
branch-and-bound algorithm is 668, while the average benefit using the sequential method is
INTELLENG
AL
577. This is not surprising as the sequential method considers only single source assignments per
ARTIFICI
target. Other significant improvements using the branch- and-bound method in the following
parameter sets are summarized below:
6. Conclusions
The primary purpose of this study was to investigate the suitability and performance of the
successive auction algorithm adopted for a class of multi-target assignment problems. The
branch and bound scheme, with its exhaustive search tree, served as the benchmark. Several
simulations were performed to compare the sequential method with the branch- and-bound
method. We found that the use of successive auctions beyond the primary source-target pairings
to generate multi-target assignments produced good results overall. While the branch-and-bound
scheme is guaranteed to find optimal results, the successive auction method consistently found
multiple assignments that are close to the optimal, with differences that may be considered
operationally insignificant. The successive auction method is extremely fast and requires fewer
enumerations. A second objective was to gain insight into the performance of the branch-and-
bound method and extend it to problems which may require multiple source assignments per
target (or target cluster). In ordering and sorting the feasible assignments by source, and
sweeping and pruning feasible solutions by target conflicts and benefit value, we found that the
first
feasible
solution
reached
NCE
INTELLENG
was a AL
ARTIFICI
good
solution
with a
benefit
close to the optimal value (that may be deemed an operationally insignificant difference) in
many cases. Thus the branch-and-bound method can be used to provide efficient solutions to the
multi- source assignment problem which may be arrived at very quickly through the use of
heuristic benefit threshold values and differences.
Military Examples: Ranger Military Robot
Features:
• Modular
• Climb Stairs
• Man Packable
• Wireless 1 mile Range

• 6-8 Hour operating time
• First Modular Robot
• Tracked Vehicle NCE

INTELLENG
AL
• Rechargeable ARTIFICI
Specifications:
Product History: Angelus Research Corp. developed the Penetrator Robot (Goldie); the first
small man packable robot introduced in 1996 for the SOMROV US Special Forces robotics
research and demonstration program. The Intruder Robot was developed as the first small All
Surface Drive off road vehicle for military service. The robust, long range, wireless control
technology developed for these robotic systems uses a dual band approach; one low
bandwidth, a 10 KHZ/channel VHF band for control and the other UHF band wide
bandwidth channel for video (encryption option is able). With this dual frequency approach
and two intelligent ARC computers used as controllers, the operator benefits by receiving
real-time control and high resolution video feeds, simultaneously.
Using dual video/audio feeds, the operator will be able to utilize true depth perception giving
the soldier the sense of being inside the vehicle as if he was an onboard driver. The control
algorithm allows other computer systems to easily control ARC robot systems over secure
communications links, and it is very simple to integrate into existing computer programs.
With only a three byte token based packet sent over the wired/wireless control link, remote
computers have total control of the robot system using differential speed control. Additional
controllers can be added for pan and tilt cameras, robotic arms, mine detection or clearing
systems, reconnaissance, hazmat sensors, etc. Multiple onboard ARC controllers can be
added for a truly multiprocessing system.
NCE
INTELLENG
AL
ARTIFICI
Modular Design:
The Ranger Modular Robot is designed to allow soldiers in the field to assemble/disassemble
the robot within minutes. The tract sections can be removed from the main body section
within two minutes time for fast assembly, disassembly, or reconfiguration. The robot can be
transported to the mission site in sections by multiple soldiers, without a single soldier giving
up his essential backpack space.

Final Document of Ai

Caricato da

Informazioni sul documento

Descrizione originale:

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Final Document of Ai

Caricato da

Copyright:

Formati disponibili

MILITRAY AND ARTIFICIAL INTELLEGENCE

resources," according to the Lockheed Martin paper. AL

• low probability of intercept/low probability of detection (LPI/LPD) waveforms;

Military Organization & Training:

Agency (DARPA) in the 1980s. Now, with a combination of military-funded development

problem is with the ambiguity of those data.

scheduled for launch later this month.

commercial practices as the bar-coded merchandise in supermarkets. A principal objective in this

Navy evaluates shipboard network to coordinate

OWENS MILLS, Md. — A form of artificial intelligence is coordinating scheduling information

NASA robot in form of snake planned to penetrate inaccessible areas NCE

Military Applications of AI:

Nowadays, developers seem to be concentrating on smaller goals, such as voice recogition

Sensor Training Data Test Data

Radar 2.5% 8.3%

Imaging 2.0% 8.0%

Fused 0.0% 4.2%

Morality: A Quick Thought:

The Generalized Weapon Target Assignment Problem

Figure1. A sample task for source 1

Then the problem may be formulated as AL

4. Branch and Bound Algorithm

Asymmetric assignment problem:

7. Extension of branch-and-bound method to multiple source assignments per target In section 4

Military Examples: Ranger Military Robot

• Wireless 1 mile Range

• 6-8 Hour operating time

• First Modular Robot

• Tracked Vehicle NCE

Potrebbero piacerti anche