NewVectors Projects
Natural Agents for Autonomous Adaptation to Threats in Varied Environments (NA3TIVE)
As the military adjusts to an increasingly wider range of mission types, simulation is proving crucial to understanding and preparing for these operations. In addition to traditional theater level operations, we need to be able to simulate military operations in urban terrain, counter terrorist activities, and futuristic operations involving large numbers of robots and agents in addition to people.
The current state of the art in simulation is inadequate to address the needs of the military. Realistic behaviors are possible for a small number of entities at the cost of expensive setup and long execution times. A large number of entities with inexpensive setup and fast execution can be simulated, but lack realistic behaviors. The challenge is to provide the best of both worlds: realistic behavior for a large number of entities with inexpensive setup and execution.
NewVectors is using agent technology for modeling real-world problems such as combat, monitoring the operation of real systems, and managing them by closing the Command and Control (C2) loop. Much of their strength is that their interactions capture the dynamics of a real system much more accurately than do analytical models.
However, a major drawback to using agents is the cost to construct and tune a full population of individual agents, particularly coarse-grained cognitive (e.g., Belief-Desire-Intention) agents. Work in swarm intelligence suggests that much of the functionality of an agent-based system can be obtained by using much simpler agents modeled on principles of artificial life. A full set of such agents can be tuned automatically using evolutionary and particle swarm techniques, providing a foundation model that provides on the order of 70% of the required functionality. Then, selected agents of particular interest can be refined into higher-level agents to support cognitive requirements such as cognitive modeling and explanation generation.
NewVectors' Department of Defense (DoD) program for Natural Agents for Autonomous Adaptation to Threats in Varied Environments (NA3TIVE), in cooperation with Soar Technology and the University of Michigan, demonstrates the feasibility of hybrid reasoning with two radically different agent architectures. High-level cognitive agents have internal structures that model cognitive concepts such as beliefs, desires, and intentions, and derive from a long research tradition in artificial intelligence. Low-level behavioral agents, inspired by research in artificial life, exhibit outward behavior that mimics domain entities, but their internal structures do not use cognitive constructs. Instead, they may use neural networks, polynomial equations, or other non-symbolic representations and processes. These two species of agents offer complementary strengths and liabilities. By integrating them in a single application, we show that it is possible to combine their respective strengths.
The demonstration context was a TacAir-Soar scenario involving guiding a mission package around Surface-to-Air Missiles (SAM) to a target. TacAir-Soar is a high fidelity human behavior model of a military pilot. In this demonstration, the cognitive agents communicated the parameters of the required path to the lightweight behavioral agents, which rapidly determined routing information for the aircraft pilots to ingress and egress a target in the face of a complex air defense, using methods inspired by insect pheromones as developed under the Joint Forces Air Component Command/Adaptive Control of Distributed Agents through Pheromone Techniques and Interactive Visualization (JFACC ADAPTIV) program. The behavioral agents then communicated the route back to the cognitive agents, which used this routing information to fly the missions to attack the designated targets.