The prime aim of US third offset strategy, announced in November 2014, was to develop effective strateges against multiple potential competitors, from small regional states like North Korea and Iran, to large advanced states like Russia and China, to non-state adversaries and actors with advanced capabilities. In response to third offset strategy DARPA has launched several programs to enhance human-machine collaboration technologies apart from A-Teams such as Resilient Synchronized Planning and Assessment for the Contested Environment (RSPACE), Collaborative Operations in Denied Environment (CODE), Distributed Battle Management (DBM), System of Systems Integration Technology and Experimentation (SoSITE), and others.
Modern military operations are dynamic and complex—requiring, for example, that infantry squads carry out their missions simultaneously in the 3-dimensional physical world, the cyber domain, and across the electromagnetic spectrum. Similarly, complexity and tempo are challenges to team operations in other applications including scientific and drug discovery, software engineering, logistics planning, advanced hardware engineering, and intelligence forecasting. Problem solving in these complex environments exceeds the capability of any individual operator and is addressed using teams. Simultaneously, autonomy and artificial intelligence are increasingly being utilized to augment and automate operations executed by teams in these environments such as command and control/battle management or cyber/electronic warfare operations.
US DOD is considering Human-machine collaboration for these complex and dynamic environments. Human-machine collaboration, specifically the ways machines can help humans with decision-making.Deputy Secretary of Defense Bob Work said that these teams up human insight with the tactical acuity of computers by allowing machines to help humans make better, faster decisions. Pairing the two will combine the ability of humans to think on the fly with the quick problem-solving methods of artificial intelligence. Work pointed to the advanced helmet on the F-35 joint strike fighter, which fuses data from multiple systems into one layout for the pilot.
US military is also employing artificial intelligence to asist operations by forming manned unmanned teams that is much more than the sum of its parts. Some examples are armed drones that fly alongside manned fighters as expendable “loyal wingmen” or scout ahead of human soldiers, AI targeting systems that line up shots for human gunners, and a host of virtual staff officers that do the organizational grunt work on everything from maintenance schedules to strike planning.
But if soldiers are going to rely on AI in battle, they will need to trust that the AI is capable of performing the mission, and that they will be able to communicate with the AI to change plans as circumstances change. This is the heart of the Adaptive Distributed Allocation of Probabilistic Tasks, or ADAPT, a $1 million contract DARPA awarded to Aptima, a Woburn, Mass.-based engineering firm in 2020. To explore this kind of human-bot collaboration, Aptima is working on a testing environment that is entirely virtual and familiar to many: Minecraft. “Teams of humans that are going to be doing some work in Minecraft and getting advised by this AI model,” said Adam Fouse, Aptima’s ADAPT Program Manager.
The goal is to find a way for humans to communicate information to an AI agent that the AI itself might not have, perhaps because it lacks the sensors to perceive it, perhaps simply because it’s in a different place. Likewise, said Fouse, the AI needs to be able to “communicate back to humans in ways that are efficient and useful, so they can understand at a glance kind of the insights of the AI.”
“ADAPT will take a significant step forward in human-AI collaboration so warfighters and intelligent technology can reason and work together to make better, faster decisions than either could do on their own,” Dr. Adam Fouse, Aptima’s ADAPT program manager, said in a statement. “By learning from its human counterparts, taking into account their goals, preferences and constraints, these more informed agents can guide AI in forecasting, creating and adapting action plans as missions evolve.” In a search-and-rescue scenario, for example, these advanced AI models and agents will think through millions of possible scenarios for commanders to choose the best plan, minimizing casualties and risks.
“Humans excel at learning from one another but can only process so much incoming information. AI, on the other hand, has incredible computational abilities but needs to learn from and communicate with humans in order to be used effectively in dynamic team situations,” Fouse explained. “These combined attributes will elevate a commander’s expertise and decision-making in fast-changing, information-intensive environments so they can respond, and adapt quickly, while considering future possibilities.” ADAPT will collect and analyze huge influxes of data for humans, optimize strategy and reallocate assets on the fly, according to the release. It is linked to two other DARPA programs: Agile Teams (A-Teams), which is also under the contract with Aptima, and Artificial Social Intelligence for Successful Teams (ASIST), where Aptima is priming the evaluation effort.
In order to transform machine learning systems from tools into partners, users need to trust their machine counterpart. One component to building a trusted relationship is knowledge of a partner’s competence (an accurate insight into a partner’s skills, experience, and reliability in dynamic environments). While state-of-the-art machine learning systems can perform well when their behaviors are applied in contexts similar to their learning experiences, they are unable to communicate their task strategies, the completeness of their training relative to a given task, the factors that may influence their actions, or their likelihood to succeed under specific conditions.
DARPA launched Competency-Aware Machine Learning (CAML) program in 2019 that addresses this challenge by enabling learning systems to be aware of their own competency. Systems will have knowledge of their learned abilities, the conditions under which those abilities were learned, knowledge of their resultant task strategies, and the situations for which those strategies are applicable.
A challenge in designing human-machine systems, however, is determining how best to meld human cognitive strengths and the unique capabilities of smart machines to create intelligent teams adaptive to rapidly changing circumstances.
To address this challenge, DARPA announced the Agile Teams (A-Teams) program in 2016, which sets out to discover, test, and demonstrate predictive and generalizable mathematical methods to enable optimized design of agile hybrid teams. A-Teams seeks to fundamentally challenge the current paradigm of human-intelligent machine systems design by changing the focus from simply using machines for automation and substitution of human capacity to an integrated fabric enabling superior collective problem solving.
DARPA Competency-Aware Machine Learning (CAML)
This competency-awareness capability contributes to the goal of transforming autonomous systems from tools into trusted, collaborative partners, DARPA officials say. Competency-aware machine learning will enable machines to control their behaviors to match user expectations and enable human operators quickly and accurately to gain insight into a system’s competence in complex, time-critical, dynamic environments.
CAML contributes to improved human-machine teaming and realization of the task synergies expected of autonomous systems. By creating a fundamentally new machine learning approach, CAML will facilitate mission planning by giving human operators insight into available machine assets based on task requirements, determining the level of autonomy to be granted, and controlling behaviors to adapt for operating conditions.
Verifying a machine’s competence increasingly is unrealistic for human operators. This can be a big problem for the military, where machines often deal with high-stake decisions, and must cope with dynamic, fast-changing conditions.
CAML is a four-year program divided into a three-year research first phase, and a one-year technology-demonstration second phase. It focuses on four technology areas: self-knowledge or experience; self-knowledge of task strategies; competency-aware learning; and capability demonstrations.
Self-knowledge of Experiences will develop mechanisms for learning systems to discover conditions encountered during operation, and maintain a memory of experiences. Self-knowledge of task strategies will enable a machine learning system to analyze its task behaviors, summarize them into generalized patterns, and identify what controls its behavior. Competency-aware learning integrates component technologies into a competency-aware learning framework that is able to communicate in human-understandable statements. It will conclude with a demonstration on a proposer-provided platform.
Officials of the U.S. Defense Advanced Research Projects Agency (DARPA) announced a $4.6 million contract to SRI International in 2019 for the Competency-Aware Machine Learning (CAML) project.
Mass. Raytheon is developing a machine learning technology under a $6 million contract from the Defense Advanced Research Projects Agency (DARPA) for the Competency Aware Machine Learning program. The contract is to facilitate communication conditions where machine learning systems from CAML can communicate to human operators what they have learned about a situation, under what conditions the machines absorbed information the best, and the situations when the learned strategies are most appropriate. This could address the problem of Machine Learning systems needing human decision assistance in combat situations.
“The CAML system turns tools into partners…Once the system has developed these skills, the team will apply it to a simulated search and rescue mission. Users will create the conditions surrounding the mission, while the system will make recommendations and give users information about its competence in those particular conditions,” said Ilana Heintz, principal investigator for CAML at Raytheon BBN Technologies.
The system makes use of a process that is similar to that in a video game, offering a list of choices and identifying a goal instead of rules. It will repeatedly play the game and learn the most effective way to achieve the goal. It will also record and explain the conditions and strategies used to come up with successful outcomes. Heintz added: “People need to understand an autonomous system’s skills and limitations to trust it with critical decisions.” After the system develops these skills, it will be applied to a simulated search and rescue mission by the team.
According to DARPA, in addition to contracting with Raytheon, CAML will continue to seek expertise in machine learning, artificial intelligence, pattern recognition, knowledge representation and reasoning, autonomous system modeling, human-machine interface, and cognitive computing.
Agile teams are defined as teams that are capable of responding to changes in goals, operating environment or team interactions such that the performance of the team is unaffected. This agility may be realized by having high robustness, flexibility, and/or resilience. Broad capability is difficult to realize in very small teams while agility is difficult to realize in large organizations; for this reason, teams of more than five and less than fifty people are of primary interest in this program.
Hybrid teams are teams that utilize intelligent machine elements to augment the performance of the team. Hybrid teams that utilize intelligent machine elements to enhance agility are the primary focus of this program. Intelligent machine elements can have a variety of possible instantiations: for example, as machine agents that are capable of peer-level interaction with, and possible substitution of, the human team members for execution of the goals of the team; or an intelligent problem solving substrate that manages team member communications and capabilities at both the individual and joint level. These intelligent machine elements may utilize a variety of context-appropriate artificial intelligence approaches and not just machine learning.
“A-Teams is focused not on developing new AI technologies per se, but on developing a framework for optimizing the use of smart machines in various roles together with humans to ensure optimal human-machine teamwork for solving dynamic problems,” DARPA program manager John Paschkewitz explained in a press release. “Given an uncertain environment and fluid team structure, how does one best use combined human and machine capabilities to make wise decisions? Are there generalizable mathematical abstractions to capture the dynamic interactions of problem space, team structure, and performance? These are the kinds of questions we intend to answer in the program.” A-team outputs will include algorithms, abstractions and architectures for a machine-based “intelligent fabric” aiming to improve decision-making on the field.
Designing Agile Human-Machine Teams
The goal of the Agile Teams (A-Teams) program is to discover, test, and demonstrate predictive and generalizable mathematical abstractions for the design of agile hybrid teams. Specifically, A-Teams will develop predictive and practically implemented abstractions and algorithms encoding the dynamic co-evolution of team structure and team problem solving strategies.
Problem solving in these complex environments exceeds the capacity of any individual and is best addressed by teams of people augmented by technology, such as with computer-aided design and collaborative work tools. A-Teams seeks to facilitate a leap forward in teamwork in which more intelligent machines in the future could not only provide automated insights but also serve as decision and interaction facilitators among team members.
The program focus will be on mathematical methods for designing optimal hybrid teams of humans and intelligent machine elements that will be demonstrated and validated in dynamic and complex problem-solving contexts using experimental testbeds. Intelligent machine elements could take a variety of possible forms, including machine agents capable of peer-level interaction with human team members for executing team goals, or as an intelligent problem solving workspace that can coordinate communications and task assignment to optimize team performance.
Intended A-Teams outputs include abstractions, algorithms, and architectures for a machine-based “intelligent fabric” that would dynamically mitigate gaps in ability, improve team decision making, and accelerate realization of collective goals. The program seeks expertise in mathematics, organizational theory, operations research, planning and scheduling, cognitive science, human factors, autonomy, and citizen science (for experimental testbeds).
A-Teams results could also apply to complexity and tempo challenges to team performance in non-combat applications, such as scientific and drug discovery, software engineering, logistics planning, advanced hardware engineering, and intelligence forecasting. Problem solving in these complex environments exceeds the capacity of any individual and is best addressed by teams of people augmented by technology, such as with computer-aided design and collaborative work tools.
A-Teams seeks to facilitate a leap forward in teamwork in which more intelligent machines in the future could not only provide automated insights but also serve as decision and interaction facilitators among team members. The results of A-Teams could also be applied to enhance human-machine collaboration technologies being developed in various DARPA programs such as Resilient Synchronized Planning and Assessment for the Contested Environment (RSPACE), Collaborative Operations in Denied Environment (CODE), Distributed Battle Management (DBM), System of Systems Integration Technology and Experimentation (SoSITE), and others.
Contract to advance the effectiveness of teaming between Humans and Machines
Aptima has announced a multi-year contract award with a potential estimated total value of $3.5 million from the Defense Advanced Research and Planning Agency (DARPA) for the Agile Teams (A-Teams) program in 1987.
Human-machine partnerships are increasingly pervasive throughout many work environments, including the military, with missions drawing together Warfighters, drones, robots, and autonomous systems. However, unlike human teams that can adapt, compensate, and apply learning from one event to the next, the blended operations of humans and networked smart systems require new models to govern and guide their complex interactions and decision-making for optimal performance.
To advance this scientific foundation, Aptima and partners, Professor Julie Shah of the Massachusetts Institute of Technology (MIT), Professor Krishna Pattipati of the University of Connecticut and Sophia Speira, LLC, are developing a framework to represent, test, optimize, and train complex socio-technical systems capable of peer-level interaction, mutual learning and problem-solving.
“The goal is to create organizations that are more dynamically responsive, that can quickly adapt to changing situations, and reallocate the right abilities, tasking, and workload across complex, mixed human-machine systems,” said Daniel Serfaty, Aptima CEO and Principal Founder.
Aptima’s approach is inspired in part by thermodynamics and the concept of energy flows, modeling energy expended on tasks, and the transfer of energy and learning from one entity and instance to another. Optimal teams minimize friction and wasted energy.
“Humans are good at collaborating and learning from each other. While current machines are not, the promise of artificial intelligence (AI) is challenging this assumption” added Serfaty. “Much like a high performing sports team uses different plays and players to anticipate and counter what an opponent might do, similarly human-machine systems must be able to collaboratively adapt as mission demands change.”
References and Resources also include: