DARPA’s Commander’s Aid: From OODA to Deep Green
DID readers send us some interesting tips. The USA’s Defense Advanced Research Projects Agency handles futuristic research projects that would be too difficult for the normal defense program R&D process (one of those projects became the Internet). Now its Information Processing Technology Office (IPTO) is turning its attention to a project called “Deep Green,” which aims to provide US commanders with significantly better decision support tools in battle. According to DARPA, Deep Green will:
“…aid in battle command and commander’s visualization by creating technologies that make it easier for the commander to articulate options to consider and anticipate the possible futures that result from those options. This proactive analysis will help predict which possible futures are becoming more likely – before they occur. Given that information, the commander can make better decisions and focus planning efforts (the generation of future branches and sequels) on where they can be the most useful.”
The article below explains the vision of Deep Green, its envisioned components, and some of the challenges the program faces. It also begins to cover contracts, now that the first R&D orders are being issued…
What Is Deep Green?
“Deep Green will develop technologies to help the commander create courses of action (options), fill in details for the commander, evaluate the options, develop alternatives, and evaluate the impact of decisions on other parts of the plan. The permutations of these option sketches for all sides and forces are assembled and passed to a new kind of combat model which generates many qualitatively different possible futures. These possible futures are organized into a graph-like structure. The commander can explore the space of possible futures, conducting “what-if” drills and generating branch and sequel options. Deep Green will take information from the ongoing, current operation to estimate the likelihood that the various possible futures may occur. Using this information, Deep Green will prune futures that are becoming very improbable and ask the commander to generate options for futures that are becoming more likely. In this way, Deep Green will ensure that the commander rarely reaches a point in the operation at which he has no options. This will keep the enemy firmly inside our decision cycle.”
“…Ideally, the OO part of OODA is done many times prior to the time when the commander must decide. When the planning and execution monitoring components of Deep Green mature, the planning staff will be working with semi-automated tools to generate and analyze courses of action ahead of the operation while the command concentrates on the Decide phase. By focusing on creating options ahead of the real operation rather than repairing the plan, Deep Green will allow commanders to be proactive instead of reactive in dealing with the enemy.
DARPA is interested in the potential contribution of emerging DARPA technologies, such as those developed under the Real-Time Adversarial Intelligence & Decision Making (RAID) program or the Multi-cell and Dismounted Command and Control (M&DC2) program, to Deep Green functions. Areas where emerging DARPA technology might be brought to bear include:
- Automated course of action generation envisioned for Commander’s Associate
- Brigade-level combat modeling envisioned for Blitzkrieg.”
Deep Green’s Components
To accomplish this, Deep Green will focus on 4 functional components
The Commander’s Associate, incl. Sketch to Plan and Sketch to Decide. Automatically converts the commander’s hand-drawn sketch with accompanying speech of his intent into a Course of Action (COA) at the brigade level. Sketch to decide takes feedback from other elements and allows the commander to explore some “what ifs” and probability branches associated with them.
Blitzkrieg. Takes the commander’s decisions and generates potential battle outcomes. Note the use of the plural term. Besides being very fast (the blitz in Blitzkrieg), it is intended to generate a broad set of possible futures. These futures should be feasible, even if not expected by human users.
Crystal Ball. Receives options from Sketch to Plan. Controls the operation of Blitzkrieg in generating futures. Takes information from the ongoing operation and updates the likelihood metrics associated with possible futures; also generates two additional metrics associated with the possible futures: value/utility and flexibility. Uses those updated likelihood metrics to prune parts of the futures graph and nominate futures at which the commander should generate additional options and invokes Sketch to Plan. Finally, it identifies upcoming decision points and invokes Sketch to Decide.
“While Crystal Ball has a moderate role prior to execution, it is the backbone of the system during execution. Once the operation is underway, Crystal Ball will get information about the ongoing operation from the battle command systems, such as FBCB2, CPoF, or the publish and subscribe services (PASS) of ABCS 6.4.”
Automated Course Of Action Generation. “Leaders from the field generally do not want machine-generated courses of action. Nevertheless, under Deep Green, DARPA intends to sponsor a small set of modest efforts to generate options automatically. While DARPA envisions Deep Green as a commander-driven battle command technology, the long-term vision of Deep Green is for options to be generated by both the commander and the computer. The intent is to insert these technologies into Deep Green in follow-on phases, so that some options are generated by humans and others are generated by machines. Initially, DARPA expects the machine generation of options to be centered on making clever mutations of the human-generated options to increase the breadth of the futures generated. Successful implementations of this capability are targeted at inclusion as a sub-module in the Sketch to Plan component of Deep Green.”
Plus these tasks:
Overall integration. Of the system’s components, and also with key external systems. This role includes architecture & standards work.
Testing & Evaluation. Of the system. DARPA intends for there to be 2 tests/experiments during each phase, about 6 months apart. The 1st test each year will be an interim status assessment. The 2nd test is designed to determine whether the goals of the system have been met for that phase.
In the 2007 Broad Agency Announcement, each of these 6 components would be bid separately, with the best performer selected for each of the tasks, but the winner of the T&E contract not be selected for any other task. That was changed in the revised BAA08-09, which kept testing and evaluation separate but removed the separation between individual components.
Program and Prognosis
The project is being handled by LTC John Surdu at DARPA IPTO, and interested firms can read the entire solicitation here (PDF, 520k – thanks to our reader for forwarding this). Industry Day for Deep Green is July 23/07 at MITRE in McLean, VA (perhaps a couple of videogame firms will show?), with initial closing for proposals on August 30/07 and Final Closing on July 9/08.
There are some conceptual issues with the thrust of Deep Green, and also some interesting areas that could offer real value. Deep Green can be an aid, but must not become a crutch or a distraction; delivering on both halves of that equation has stymied just about every system of this type to date. DARPA is certainly correct that machines have dealt poorly with the kinds of “intuitively expressed plans” that are common and necessary in battle; that existing AI planning & monitoring systems have been too linear, deterministic, and automated rather than assistive; and that current combat models are slow, narrow, and need a lot of hand-holding. On a battlefield, the level of maintenance and support “tail” required also becomes an issue – and the unpredictable nature of events means that commanders must be able to use the system well while never coming to rely on it.
Can Deep Green cut all of those Gordian Knots? Maybe. Maybe not. Still, Deep Green does explore some interesting ideas – and probably lays out the map for a very successful video game interface and concept that we can expect to see in simplified form at home within the next 10-20 years. Quicker, iterative exploration of plans and what-ifs, so that commanders have been exposed to some possible outcomes beforehand and thought them through, certainly stands out as a big potential plus in the Deep Green concept.
The truth about DARPA’s projects is that very few of them are successful across the board – but they don’t require that level of success in order to be worth the time and money. If some elements from the project prove out in practice and make their way into current systems, then DARPA will be able to claim success whether or not Deep Green as a whole turns out to be viable, needs a few more iterations of Moore’s Law to make sense, or turns out to be a category error of a concept.
Contracts and Key Events
Unless it says otherwise, dates given are the US DoD’s DefenseLINK announcement dates.
June 11/09: Science Applications International Corp. (SAIC) in Science Drive, Orlando, FL received a $6.6 million cost-plus fixed fee contract for Deep Green Task 1. Task 1 encompasses development of full Deep Green systems, including the various functional components, such as the commander’s associate, crystal ball, and blitzkrieg.
Work will be performed in Orlando, FL and locations across the United States, and is expected to be complete by July 30/09. Web bids were solicited on Dec 30/07, and 11 bids were received by the CECOM Acquisition Center at Fort Monmouth, NJ (W15P7T-08-C-M011).
June 2/08: BAE Systems National Security solutions in Burlington, MA received a $5.2 million cost-plus fixed fee contract for the Defense Advanced Research Project Agency Deep Green Program for the “development of a battle command decision and support system.” Task 1 encompasses development of full Deep Green systems, including the various functional components, such as the commander’s associate, crystal ball, and blitzkrieg.
Work will be performed in locations across the United States, and is expected to be complete by July 31/09. Web bids were solicited on Dec 21/07, and 11 bids were received by the CECOM Acquisition Center at Fort Monmouth, NJ (W15P7T-08-C-M002).
May 23/08: Applied Research Associated in Orlando, FL, is picked as the test and evaluation contractor for Deep Green.
- FBO – DARPA Deep Green Solicitation
- DARPA IPTO Industry Day
- WIRED Danger Room (July 19/07) – Pentagon Plots Digital “Crystal Ball” to “See the Future” in Battle. Actually, if you read the solicitation, that’s not quite what they want. That why we read it – and why we recommend that interested readers take the time. It’s actually quite readable and understandable.
- Federal Computer Weekly (July 17/07) – DARPA wants better decision-making on battlefield