Machine learning, reasoning, perception provide cognitive assistants for novel, robust functions.
Col. Jon Krenkel, USAF, commander, U.S. Air Force Command and Control Battlelab, demonstrates the features of the master air attack plan toolkit. Software applications such as the toolkit and the Personalized Assistant that Learns program being developed at the Defense Advanced Research Projects Agency will reduce the number of senior officers required to support air tasking orders during operations.
Significant individual technology advances are being harnessed to facilitate effective cognitive computing systems. These information system technologies focus on a common application that radically improves the way computers support human beings. A cognitive system is emerging that can reason, learn from experience, be told what to do, explain its actions and respond robustly to surprise.
Known as PAL, for Personalized Assistant that Learns, the program underway at the Defense Advanced Research Projects Agency (DARPA),
Developing cognitive systems that can learn and adapt to their user would dramatically improve a wide range of military operations, according to David Gunning. He is the PAL program manager in DARPA’s Information Processing Technology Office. Funding for the program this year is approximately $35 million. Gunning’s bachelor’s degree in psychology from
“Developments in machine learning, reasoning, perception and multimodal interaction are areas of important advances that all relate to cognitive computing. Improvements in processors, memory, sensors and networking have also dramatically changed the context of cognitive systems research,” Gunning illustrates. “All these areas are coming together to focus on a common application—PAL.”
Gunning adds that current commercial and military software systems are plagued by brittleness and the inability to deal with change and novel situations. Most software must be painstakingly programmed for every contingency. If PAL succeeds, it could result in software systems that learn on their own and can adapt to changing situations without the need for constant reprogramming. PAL technology could drastically reduce the money the U.S. Defense Department spends on information systems of all kinds.
PAL is the first broad-based research program in cognitive systems since the strategic computing initiative DARPA funded in the 1980s. ”We’ve made huge increases in all of the individual technologies, especially in machine learning, where we are focusing, along with vision systems and speech understanding—products of machine learning. Technical progress in these areas leads us to take another look at integrating them into a complete system and see how close we can come to PAL in a major tiered approach,” Gunning remarks.
“Approximately half of the PAL funding is directed to some 20 universities and research centers involved in cognitive computing developments. While most of the research is in universities that are involved in state-of-the-art machine learning and reasoning, industry also is involved. Companies such as SRI International, ISI and Lockheed Martin Advanced Technology Laboratories are participating,” Gunning discloses. “While research continues, the bulk of the technology program involves system integration at organizations such as SRI. This company assembles technology advances and each year provides an integrated cognitive system.
“However, PAL is not simply attempting to assemble technologies but to create new techniques in reasoning and learning for tightly coupled, more powerful systems. And we are just beginning to apply this approach to create an intelligent assistant to perform many Defense Department applications. Areas include the Army’s Command Post of the Future, the U.S. Navy’s Composeable FORCEnet and the U.S. Air Force’s Air Tasking Order [ATO],” Gunning continues. Developing intelligent systems to support decision-making may provide dramatic advances for traditional military roles and missions. Technologies developed under the PAL program are intended to make military decision-making more efficient and effective at all levels.
SRI is working with 20 subcontractors and is the prime contractor in two major efforts. One involves a program called Cognitive Assistant that Learns and Organizes, or CALO, and the other is Reflective Agents with Distributed Adaptive Reasoning, or RADAR. Military commanders and staff work in dynamic environments, so CALO must be able to adapt to this environment, learning new concepts, tactics and tasks through observation and inference. User advice and instruction, both implicit and explicit, are added. This learning must occur “in the wild,” meaning that the system learns through its exposure to the environment, without intervention from human programmers.
Learning in the wild is a radical research approach that requires significant rethinking of most methods in the machine learning field, often making them online and interactive, and always embedding them in the system’s reasoning. User interface components also impact system knowledge and functions. In turn, this approach requires rethinking the reasoning, user interface and knowledge representation technology, Gunning states.
The RADAR project is building and empirically evaluating a cognitive assistant that learns to help a human user in situations of intense information overload. The specific kind of information overload considered by RADAR is a flood of inbound e-mail messages in a crisis situation. As part of the overall PAL program, RADAR’s overriding theme also is learning in the wild. As such, RADAR must be usable by a person with no special training, and it must learn during normal use.
The DARPA Command Post of the Future is a fielded command and control software system in use by the U.S. Army and U.S. Marine Corps in
Any military command or intelligence center where there is a flood of information is an area targeted by PAL to package intelligent software in a way that provides an assistant for the commander, intelligence officer or staff member. “This software assistant would sit behind the scenes, watch the user and observe the information that person needs, and then retrieve similar information—learning how to perform those tasks and how to make related suggestions,” Gunning points out.
A PAL prototype is being built for use with the U.S. Strategic Command’s SKIWeb information system beginning in July, and a spin-off also will be used this summer with an Army Knowledge Online system called Company Commander.com. The technology also is with the Air Force Air Combat Command and the ATO coordinator.
The ATO development process requires 72 hours. Three ATOs are in progress at any time, and a new one is released every 24 hours. Numerous steps are involved and must be coordinated to ensure that ATO objectives are met. The goal of PAL is to develop an assistant that interacts electronically with the Air Force Theater Battle Management system to provide data support and communications. This task currently requires 12 senior officers operating in multiple shifts on ATOs
The Navy is developing FORCEnet as its supporting component to the Net-Enabled Command Capability, the joint program that is being implemented as a services-oriented command, control, communications, computers, intelligence, surveillance and reconnaissance capability. The PAL program is identifying, implementing and demonstrating planning and information management technologies. As part of this effort, PAL will assist in generating information blogs to enhance situational awareness.
PAL is developing a series of prototype cognitive systems that can act as an assistant for commanders. DARPA will get its own version of a PAL system this summer to evaluate in its offices. The program’s success would usher in a new era of computational support for a broad range of human activity, especially military functions.
DARPA Personalized Assistant that Learns: www.darpa.mil/ipto/Programs/pal/index.htm