s o s: DARPA's Raiding the (Enemy’s) Mind

PreCrime Detector?: Brain scanner predicts your future moves

RAID Reading the mind of your enemy may soon become easier. Computers may be able to do so even better than humans, experiments conducted by the Defense Advanced Projects Research Agency (DARPA) Real-time Adversarial Intelligence and Decision-making (RAID) program suggest.

RAID is a tool for semi-automated generation of enemy estimates. Its job is to anticipate the upcoming actions of the enemy, and do so not just before, but also during the unfolding battle, in near real-time. In a way, one may say the purpose of RAID is to read the mind of the enemy.

In less flamboyant terms, RAID performs what some call enemy estimates in mission analysis and preparation, or running estimates during the execution of a mission. Others see in such a capability elements of Level 3 information fusion and data mining.

To stress this emerging capability, the RAID program focuses on a particularly challenging and currently relevant environment: a fluid urban fight against a dismounted insurgent force.

Based on multiple simulation-based experiments as well as on a live-force demonstration, DARPA researchers find that the RAID software can, on average, be at least comparable to a competent staff in terms of accuracy, and may offer significant advantages in terms of speed and personnel requirements.

DARPA and the Army are currently working together to transition RAID into the Army Distributed Common Ground System-Army (DCGS-A) program of record.

In developing RAID, DARPA responds to a well-recognized unmet need. Reading the enemy mind is hard. According to Colonel Joe Moore (Ret.), who commanded the OPFOR at the Army National Training Center (NTC) in 2001-2005 and now serves as one of RAID’s military advisers, “At NTC, the BLUFOR can rarely predict OPFOR positions or intent, in either conventional or insurgency scenarios.”

Reflecting such concerns, the DCGS-A Operational Requirements Document calls for DCGS-A to include the means to estimate the “near-future enemy positions and actions at intervals.” Similarly, the Army TRADOC Force Operating Capabilities pamphlet stated that the Army needs tools “for performance of semi-automated predictive analysis.”

Not everyone is comfortable, however, with the air of certainty implicit in the term “predictive.” DARPA RAID researchers acknowledge that infallible predictions are nowhere near on the horizon, and that neither humans nor computers can predict the actions of an intelligent adversary with anything approaching certainty.

However, in the spirit of the famous Turing test for artificial intelligence, it is perfectly reasonable to aim at a more realistic objective for a computational “mind-reader”: rapidly generating enemy estimates that are at least comparable in quality and usefulness to those produced by competent humans.

20050807FMG33969.wmv

Free-Play Wargames

Since its inception in 2004, the DARPA RAID program has pursued a rigorous strategy of experiment-based development. Roughly every six months, each iteration of the RAID software has undergone an extensive experimental evaluation.

For example, a series of experiments included nearly two dozen simulation-based, free-play wargames pitting a Blue dismounted company, with five Stryker vehicles, in an urban fight against a well prepared force of hostile militias and insurgent cadres. Unfolding in a capital of a developing country, the battles were partially patterned after actual events and missions executed by the U.S. military in recent years and months.

Commanding the simulated insurgent fighters within a customized version of the Army OneSAF Testbed system, the experienced Red team played an intelligent, innovative, hard-to-template adversary. Red’s creative use of urban terrain, rapid movement in the familiar city, concealment, deceptions, ambushes, IEDs, RPGs, heavy machine guns, infiltration and civilian spies challenged the simulated Blue forces and their human commanders.

As each wargame progressed, RAID continually generated a series of running estimates—enemy goals, strengths, positions, movements—looking forward as far as 10 minutes to two hours. Independently, another set of predictive running estimates was produced by a competent military staff with relevant operational experience.

Remarkably, when the two sets of predictions were compared to what the Red force actually did in the simulated battles, RAID predictions were comparable—and occasionally even significantly more accurate—to staff’s. In particular, RAID was often surprisingly accurate in pinpointing the likely locations of concealed insurgent teams and estimating their future re-positioning.

With emerging innovative capabilities like RAID, defense technologists always look for opportunity to experiment in controlled and accessible yet realistic environments, preferably with live forces. Yet there are remarkably few such experimental venues. One of them is the series of experiments conducted by the product manager, C4ISR on-the-move (OTM), an organization within Army CERDEC, headquartered in Fort Monmouth, N.J. When DARPA wanted to try RAID in a live-force experiment, the PM C4ISR OTM, Lieutenant Colonel William T. Utroska, offered them the opportunity to participate in the August 2007 experiment at Fort Dix, N.J.

Although RAID had been tested in multiple wargaming experiments, the 2007 live-force experiment at Fort Dix was an outstanding opportunity to test the system with real soldiers and realistic systems. The company commander and his subordinates used RAID in a C2 vehicle, accessing it via the Force XXI Battle Command Brigade and Below (FBCB2) system. It was also used by the battalion staff at a workstation in the battalion TOC. RAID estimates were also made available on FBCB2 display to a platoon leader.

Before and during the mission, RAID continuously read information from FBCB2, such as enemy spot reports and friendly positions. It also received input and assumptions from the battalion S2 or the company commander, using an experimental interface within FBCB2. RAID can also read and reason on voluminous data like SIGACTs (significant activity reports) in the region if they are available from appropriate databases

Then, combining all such data, RAID generated running estimates of what the enemy was likely to do, and where the enemy was currently located, including likely locations of IEDs, ambushes, concealed enemy positions, and their routes of infiltration or retreat.

During a typical multi-hour mission, responding to dozens of requests from the personnel riding in the C2 vehicle, RAID would rapidly produce estimates of the enemy current situation and intent. Upon the completion of the experiment, the uniformed personnel provided the RAID researchers with encouraging feedback and suggestions for additional features.

RAID’s performance at the C4ISR OTM experiment helped spur the Army’s interest in additional applications of this technology. According to Stilman Advanced Strategies, one of RAID’s developers, the Army awarded the company a contract for a further effort to explore means to integrate RAID with the Army FBCB2 system, and to demonstrate the technology this fall in an Army experiment called the Air Assault Expeditionary Force (AAEF).

DCGS-A Transition

Although integration of RAID capabilities into a future version of FBCB2 is an important possibility, a more immediate application of RAID is within the Army DCGS-A system. Current plans are to integrate RAID technologies into one of the upcoming versions of DCGS-A, and to assess its value to warfighters under variety of operational conditions.

To this end, in early 2007 DARPA and DCGS-A jointly developed an initial experimental integration of RAID with elements of the DCGS-A system, and performed an experimental evaluation that involved operationally relevant data and scenarios. The experiment confirmed the technical feasibility of the integrated concept and provided basis for additional modifications and for design of further evaluations.

In particular, there is a clear need to assimilate RAID technologies into the service-oriented architecture (SOA) of DCGS-A. “RAID combines several interesting technologies, each with its own valuable outputs. So, rather than integrate a monolithic application into DCGS-A, we look for ways to turn each such sub-technology into a service,” said Alan Hansen, director of the DCGS-A System Integration Lab.

“For example, one of the RAID technologies generates estimates of IED threats,” Hansen explained. “In a service-oriented architecture, the technology can operate as service, and if a human user or another system needs IED-related estimates, it can request such estimates directly from this particular service.”

The SOA-based integration of RAID into DCGS-A is currently underway, to be followed by further evaluations.

Other Army programs and organizations are exploring additional applications of RAID technologies. For example, Soar Technologies, another RAID developer, recently announced an Army contract to integrate RAID with the Collection Management Tool (CMT) developed and deployed by the Army Battle Command Battle Lab at Fort Huachuca, Ariz. In this application, RAID would help collection planners and manager by providing a continuous, dynamic intelligence running estimate of the battlefield, a map of battlefield uncertainty and threat over time, and annotated search locations targeted specifically at reducing uncertainty in high threat regions.

Like any technologies, weapons or tools, RAID can be useful only if applied properly. Military advisers to DARPA see great opportunities in RAID, but also caution about the need for appropriate use.

“Frankly I was skeptical about RAID early in the program,” said Major General Waldo Freeman (Ret.), a combat-experienced infantry officer who advised DARPA on the RAID program. “But after watching it mature for three years, I have come to appreciate it as a potentially powerful tool. RAID already offers the tactical user numerous options for its use, and they will invent more as they learn to appreciate its capabilities. Most importantly, it helps stimulate the human cognitive process, and helps commanders under pressure think about the tactical problem at hand. Used properly it will help produce better decisions.”

“The implications of RAID are enormous,” said Brigadier General Wayne M. Hall (Ret.). “It can encourage people to think aggressively and creatively about the operational environment and what a smart, adaptive foe could be doing. It also provides automated assistance to act/react/counteract wargaming, and the means to mitigate risk. The wise commander and his intelligence officer can use RAID-estimated enemy actions to focus ISR and provide warning if indeed the RAID hypothesis is coming into fruition.”

“In my experience four decades ago both in the field in Germany and in war zone D in Vietnam, as a company commander at night I often planned under a poncho with a flashlight,” Freeman recalled. “I spent virtually all my effort on movement or positioning my own platoons and weapons because my knowledge of the opponent was so fuzzy. Well, we have come a long way!”

Dr. Alexander Kott is a program manager at DARPA’s Information Exploitation Office.

Real-Time Adversarial Intelligence & Decision Making (RAID)

Mission:
The RAID Program is developing novel capabilities for anticipating enemy actions and deceptions, with particular focus on providing real-time support to tactical commanders in urban operations. Leveraging emerging technologies in adversarial and deception reasoning, the Real-time Adversarial Intelligence and Decision-making (RAID) program intends to address the difficult technical problems to support the U.S. Army and Air Force in their need for tools for predictive analysis and predictive battlespace awareness.

Problem Statement:

Provide predictive, anticipative analysis of enemy future actions while making effective assumptions and suggestions for friendly actions
Identify enemy's attempts to conceal its assets and actions and to deceive the friendly forces
Monitor the unfolding operation and continuously confirm, disconfirm and update the products of predictive analysis
Enable the predictive analysis support in a transparent fashion that does not impose additional workload on the commander and staff

Technical Objectives:

Prove that adversarial reasoning can be automated and can generate high quality predictions of enemy actions
Prove that automated reasoning can be robust and effective in the presence of concealment, deception, and the impact of doctrinal and cultural biases
Integrate the predictive analysis tools into a warfighter's C2 and intelligence support system, Army DCGS-A.

Vision:
The RAID Program focuses on the challenge of anticipating enemy actions in a military operation. In a number of recent publications, US military leaders call for development of techniques and tools to address this critical challenge.

The US Air Force community uses the term predictive battlespace awareness while a related term, predictive analysis, is beginning to be used in the US Army community. Both refer to future techniques and technologies that would help the commander and staff to characterize and predict likely enemy courses of action, to relate the history of the enemy's performance to its current and future actions, and to associate these predictions with opportunities for friendly actions and effects.

Both communities have pointed out the lack of technologies, techniques and tools to support predictive analysis and predictive battlespace awareness.

The RAID program will result in key technologies for tools capable of in-execution predictive analysis of enemy probable actions. A particular focus of the program will be tactical urban operations against irregular combatants - an especially challenging and operationally relevant domain.

The program intends to leverage novel approximate game-theoretic and deception-sensitive algorithms to provide real-time enemy estimates to tactical commander. In doing so, RAID will address two critical technical challenges: (a) Adversarial Reasoning: the ability to continuously identify and update predictions of likely enemy actions; (b) Deception Reasoning: the ability to continuously detect likely deceptions in the available battlefield information.

Realistic experimentation and evaluation will drive the development process using human-in-the-loop, Army OneSAF-based wargames to compare humans and RAID.

The products of the program are targeted for transition to the Army, such as the DCGS-A program.

Goals:

The RAID program will be conducted in three 12-month phases. Funding for later phases is entirely contingent upon meeting system-level performance goals established for earlier phases.

Phase I - Adversarial Anticipation And Counteraction: Develop mechanisms to compute adversarial, anticipative, move-countermove actions.
Phase II - Adversarial Reasoning About Concealment And Deception: Develop ability to see through fog or war and recognize deceptions.
Phase III - Integration and Transition to Army DCGS-A: Develop fieldable products which can integrate with existing C2 and ISR systems.

PHASE I, II, AND III GATE CRITERIA:

Experiments at the end of each phase will yield several key metrics to be used for the go-nogo decisions regarding the continuing funding of the RAID program. All metrics assume the experimental conditions summarized in the Experimental Plan.

Power: the computational scale of the problem that RAID can solve, measured by the nominal search space of the problem. This number grows with increased number and types of assets on Blue and Red side, complexity of terrain, increased number and granularity of the combatants' actions, etc. For comparison, this number for the game of chess is typically estimated at 10**35. RAID is looking at astronomically larger problems. Larger is better. PI > 10**8,000; PII > 10**20,000; PIII > 10**60,000]

Speed: time required by RAID in order to deliver its predictive estimates to the user; measured from the moment when the user requests RAID estimated and until the estimates are displayed on the user interface. Smaller number (faster) is better. [PI <>

Workload: the minimal number of people (full-time equivalents) necessary to operate the RAID system during the execution of an operation. Ideally, the user-operator of the RAID tool will need only a fraction of his time to attend to the needs of the tool. Smaller is better. [PI = 2; PII = 1; PIII = 0.5]

Accuracy: the number of wrong predictions made by RAID, expressed as a fraction of total predictions and compared statistically to the same measure of human expert performance. Typical predictions will refer to tangible estimates used in the practice of Military Intelligence, such as location, strengths and actions of an enemy unit at a particular time interval in the future. Wrong predictions include false positive - red actions that are predicted but do not occur, false negative - red actions occur but are not predicted. This measure will not be used in Phase 1 due to the fact that the focus of development in Phase 1 will be on predicting what Army doctrine calls most dangerous actions as opposed to most likely (influenced by Red culture, etc.) which will be a focus of Phase 2. Lower is better. [PI = n/a; PII „T 1.0; PIII „T 1.0]

Effectiveness: the overall score (combining such factors as advance to objectives, destruction of the Red force, etc.) achieved by the Blue side. The score achieved in the test series (with RAID and 1-2 personnel) will be compared statistically to those achieved by a full staff of human experts (5-7 personnel) without RAID. Higher is better. [PI „d 1.0; PII „d 1.0; PIII „d 1.0]

Technical Challenges:

Tight interdependence, coupling of blue and red actions.
Blue knowledge of red assets and actions is inevitably limited. Observations as well as interpretations of the observations are subject to a significant degree of errors and latency.
In addition to partial, delayed and often erroneous observations, the battlefield knowledge is limited by a purposeful, continuous, aggressive, intelligent concealment and deception.
Cultural, doctrinal, psychological effects. It is not enough to consider the most dangerous course of action. The most likely course of action can be significantly different from the theoretically most advantageous one.
Complex urban terrain offers a high density of threats and opportunities for forces.
Further, the terrain itself is dynamic because it is modified by human actions.
The presence of non-combatants on the battlefield must be explicitly considered.
Fire and maneuver of forces are not the only actions that must be carefully considered. Intelligence gathering, communications, and logistics (including casualty evacuation) are tightly coupled with fire and maneuver.
The scale of the computational problem is immense and yet solutions must be generated in near real-time.
To be of practical value, a successful technical approach must allow for easy modification and extension of the coverage.

Specific Technical Challenges:

Adversarial Reasoning: continuously identify and update predictions of likely enemy actions
Deception Reasoning: continuously detect likely deceptions in the available battlefield information

Associated Attachment(s)