Have you ever been told that all you do is play games , thus
you are useless in real world? Then check this out.
Artificial intelligence (AI)
developed by a University of Cincinnati doctoral graduate was recently assessed
by subject-matter expert and retired United States Air Force Colonel Gene Lee -
who holds extensive aerial combat experience as an instructor and Air Battle
Manager with considerable fighter aircraft expertise - in a high-fidelity air
combat simulator.
The artificial intelligence, dubbed
ALPHA, was the victor in that simulated scenario, and according to Lee, is
"the most aggressive, responsive, dynamic and credible AI I've seen to
date."
Details on ALPHA - a significant
breakthrough in the application of what's called genetic-fuzzy systems are
published in the most-recent issue of the Journal of Defense Management,
as this application is specifically designed for use with Unmanned Combat
Aerial Vehicles (UCAVs) in simulated air-combat missions for research purposes.
The tools used to create ALPHA as
well as the ALPHA project have been developed by Psibernetix, Inc., recently
founded by UC College of Engineering and Applied Science 2015 doctoral graduate
Nick Ernest, now president and CEO of the firm; as well as David Carroll,
programming lead, Psibernetix, Inc.; with supporting technologies and research
from Gene Lee; Kelly Cohen, UC aerospace professor; Tim Arnett, UC aerospace
doctoral student; and Air Force Research Laboratory sponsors.
High pressure and fast pace: An
artificial intelligence sparring partner
ALPHA is currently viewed as a
research tool for manned and unmanned teaming in a simulation environment. In
its earliest iterations, ALPHA consistently outperformed a baseline computer
program previously used by the Air Force Research Lab for research. In other
words, it defeated other AI opponents.
In fact, it was only after early
iterations of ALPHA bested other computer program opponents that Lee then took
to
In fact, it was only after early
iterations of ALPHA bested other computer program opponents that Lee then took
to manual controls against a more mature version of ALPHA last October. Not
only was Lee not able to score a kill against ALPHA after repeated attempts, he
was shot out of the air every time during protracted engagements in the
simulator.
Since that first human vs. ALPHA
encounter in the simulator, this AI has repeatedly bested other experts as
well, and is even able to win out against these human experts when its (the
ALPHA-controlled) aircraft are deliberately handicapped in terms of speed,
turning, missile capability and sensors.
Lee, who has been flying in
simulators against AI opponents since the early 1980s, said of that first
encounter against ALPHA, "I was surprised at how aware and reactive it
was. It seemed to be aware of my intentions and reacting instantly to my
changes in flight and my missile deployment. It knew how to defeat the shot I
was taking. It moved instantly between defensive and offensive actions as
needed."
He added that with most AIs,
"an experienced pilot can beat up on it (the AI) if you know what you're
doing. Sure, you might have gotten shot down once in a while by an AI program
when you, as a pilot, were trying something new, but, until now, an AI opponent
simply could not keep up with anything like the real pressure and pace of
combat-like scenarios."
But, now, it's been Lee, who has
trained with thousands of U.S. Air Force pilots, flown in several fighter
aircraft and graduated from the U.S. Fighter Weapons School (the equivalent of
earning an advanced degree in air combat tactics and strategy), as well as
other pilots who have been feeling pressured by ALPHA.
And, anymore, when Lee flies against
ALPHA in hours-long sessions that mimic real missions, "I go home feeling
washed out. I'm tired, drained and mentally exhausted. This may be artificial
intelligence, but it represents a real challenge."
An artificial intelligence wingman:
How an AI combat role might develop
Explained Ernest, "ALPHA is
already a deadly opponent to face in these simulated environments. The goal is
to continue developing ALPHA, to push and extend its capabilities, and perform
additional testing against other trained pilots. Fidelity also needs to be
increased, which will come in the form of even more realistic aerodynamic and
sensor models. ALPHA is fully able to accommodate these additions, and we at
Psibernetix look forward to continuing development."
In the long term, teaming artificial
intelligence with U.S. air capabilities will represent a revolutionary leap.
Air combat as it is performed today by human pilots is a highly dynamic
application of aerospace physics, skill, art, and intuition to maneuver a
fighter aircraft and missiles against adversaries, all moving at very high
speeds. After all, today's fighters close in on each other at speeds in excess
of 1,500 miles per hour while flying at altitudes above 40,000 feet. Microseconds
matter, and the cost for a mistake is very high.
Eventually, ALPHA aims to lessen the
likelihood of mistakes since its operations already occur significantly faster
than do those of other language-based consumer product programming. In fact,
ALPHA can take in the entirety of sensor data, organize it, create a complete
mapping of a combat scenario and make or change combat decisions for a flight
of four fighter aircraft in less than a millisecond. Basically, the AI is so
fast that it could consider and coordinate the best tactical plan and precise
responses, within a dynamic environment, over 250 times faster than ALPHA's
human opponents could blink.
So it's likely that future air
combat, requiring reaction times that surpass human capabilities, will integrate
AI wingmen - Unmanned Combat Aerial Vehicles (UCAVs) - capable of performing
air combat and teamed with manned aircraft wherein an onboard battle management
system would be able to process situational awareness, determine reactions,
select tactics, manage weapons use and more. So, AI like ALPHA could
simultaneously evade dozens of hostile missiles, take accurate shots at
multiple targets, coordinate actions of squad mates, and record and learn from
observations of enemy tactics and capabilities.
UC's Cohen added, "ALPHA would
be an extremely easy AI to cooperate with and have as a teammate. ALPHA could
continuously determine the optimal ways to perform tasks commanded by its
manned wingman, as well as provide tactical and situational advice to the rest
of its flight."
A programming victory: Low computing
power, high-performance results
It would normally be expected that
an artificial intelligence with the learning and performance capabilities of ALPHA,
applicable to incredibly complex problems, would require a super computer in
order to operate.
However, ALPHA and its algorithms
require no more than the computing power available in a low-budget PC in order
to run in real time and quickly react and respond to uncertainty and random
events or scenarios.
According to a lead engineer for
autonomy at AFRL, "ALPHA shows incredible potential, with a combination of
high performance and low computational cost that is a critical enabling
capability for complex coordinated operations by teams of unmanned aircraft.
Ernest began working with UC
engineering faculty member Cohen to resolve that computing-power challenge
about three years ago while a doctoral student. (Ernest also earned his UC
undergraduate degree in aerospace engineering and engineering mechanics in 2011
and his UC master's, also in aerospace engineering and engineering mechanics,
in 2012.)
They tackled the problem using
language-based control (vs. numeric based) and using what's called a
"Genetic Fuzzy Tree" (GFT) system, a subtype of what's known as fuzzy
logic algorithms.
States UC's Cohen, "Genetic
fuzzy systems have been shown to have high performance, and a problem with four
or five inputs can be solved handily. However, boost that to a hundred inputs,
and no computing system on planet Earth could currently solve the processing
challenge involved - unless that challenge and all those inputs are broken down
into a cascade of sub decisions."
That's where the Genetic Fuzzy Tree
system and Cohen and Ernest's years' worth of work come in.
According to Ernest, "The
easiest way I can describe the Genetic Fuzzy Tree system is that it's more like
how humans approach problems. Take for example a football receiver evaluating
how to adjust what he does based upon the cornerback covering him. The receiver
doesn't think to himself: 'During this season, this cornerback covering me has
had three interceptions, 12 average return yards after interceptions, two forced
fumbles, a 4.35 second 40-yard dash, 73 tackles, 14 assisted tackles, only one
pass interference, and five passes defended, is 28 years old, and it's
currently 12 minutes into the third quarter, and he has seen exactly 8 minutes
and 25.3 seconds of playtime.'"
That receiver - rather than standing
still on the line of scrimmage before the play trying to remember all of the
different specific statistics and what they mean individually and combined to
how he should change his performance - would just consider the cornerback as
'really good.'
The cornerback's historic capability
wouldn't be the only variable. Specifically, his relative height and relative
speed should likely be considered as well. So, the receiver's control decision
might be as fast and simple as: 'This cornerback is really good, a lot taller
than me, but I am faster.'
At the very basic level, that's the
concept involved in terms of the distributed computing power that's the
foundation of a Genetic Fuzzy Tree system wherein, otherwise, scenarios/decision
making would require too high a number of rules if done by a single controller.
Added Ernest, "Only considering
the relevant variables for each sub-decision is key for us to complete complex
tasks as humans. So, it makes sense to have the AI do the same thing."
In this case, the programming
involved breaking up the complex challenges and problems represented in aerial
fighter deployment into many sub-decisions, thereby significantly reducing the
required "space" or burden for good solutions. The branches or sub
divisions of this decision-making tree consists of high-level tactics, firing,
evasion and defensiveness.
That's the "tree" part of
the term "Genetic Fuzzy Tree" system.
Programming that's language based,
genetic and generational
Most AI programming uses
numeric-based control and provides very precise parameters for operations. In
other words, there's not a lot of leeway for any improvement or contextual
decision making on the part of the programming.
The AI algorithms that Ernest and
his team ultimately developed are language based, with if/then scenarios and
rules able to encompass hundreds to thousands of variables. This language-based
control or fuzzy logic, while much less about complex mathematics, can be
verified and validated.
Another benefit of this linguistic
control is the ease in which expert knowledge can be imparted to the system.
For instance, Lee worked with Psibernetix to provide tactical and
maneuverability advice which was directly plugged in to ALPHA. (That
"plugging in" occurs via inputs into a fuzzy logic controller. Those
inputs consist of defined terms, e.g., close vs. far in distance to a target;
if/then rules related to the terms; and inputs of other rules or
specifications.)
Finally, the ALPHA programming is
generational. It can be improved from one generation to the next, from one
version to the next. In fact, the current version of ALPHA is only that - the
current version. Subsequent versions are expected to perform significantly
better.
Again, from UC's Cohen, "In a
lot of ways, it's no different than when air combat began in World War I. At
first, there were a whole bunch of pilots. Those who survived to the end of the
war were the aces. Only in this case, we're talking about code."
To reach its current performance
level, ALPHA's training has occurred on a $500 consumer-grade PC. This training
process started with numerous and random versions of ALPHA. These automatically
generated versions of ALPHA proved themselves against a manually tuned version
of ALPHA. The successful strings of code are then "bred" with each
other, favoring the stronger, or highest performance versions. In other words,
only the best-performing code is used in subsequent generations. Eventually,
one version of ALPHA rises to the top in terms of performance, and that's the
one that is utilized.
This is the "genetic" part
of the "Genetic Fuzzy Tree" system.
Said Cohen, "All of these
aspects are combined, the tree cascade, the language-based programming and the
generations. In terms of emulating human reasoning, I feel this is to unmanned
aerial vehicles what the IBM/Deep Blue vs. Kasparov was to chess."
EmoticonEmoticon