Adaptive Distributed Allocation of Probabilistic Tasks (ADAPT II)

The future success of military teams operating in dynamic and uncertain environments will require the incorporation of artificial intelligence (AI) to help structure those teams, create plans of actions, execute those plans, and adapt plans as the environment and goals change. Successfully combining human of AI team members can achieve better results than either could on their own, but an uninformed approach to deployment of AI could reduce rather than improve effectiveness (Woods and Hollnagel, 2006; Schurr et al, 2019). Achieving successful incorporation of AI into a team requires collaborative AI that perceives, learns, and adapts based on observations of the environment, interactions with human team members, and simulations of possible future states. However, current state-of-the-art AI agents lack the following critical capabilities needed to accomplish this vision: (1) cognitive skills, including exhibiting human-like curiosity, biases, and errors; (2) ability to learn complex tasks quickly with limited feedback; and 3) ability to coordinate and co-learn with human or AI teammates (Demir et al, 2017).   In response to this need, Aptima and Arizona State University (ASU) propose to develop Adaptive Distributed Allocation of Probabilistic Tasks (ADAPT) to design, build, and validate collaborative, adaptive, and realistic AI agents. These agents will interact with team members to guide them in planning and execution of missions, selection of action, and adaptation to changing world dynamics. Our model formalism is based on the biologically plausible theory developed in computational neuroscience and psychology called active inference (Friston, 2010), which Aptima has applied to model behaviors of multi-agent teams (Levchuk et al., 2017, 2018). This approach will enable ADAPT agents to execute four processes fundamental to human cognition: learning, perception, control/planning, and simulation. Through these processes, ADAPT agents will create plans and predictions that jointly minimize a variational free energy, which is a bound on uncertainty of the agents about their past and future outcomes. ADAPT-generated behaviors will simultaneously maximize extrinsic and intrinsic rewards, resulting in efficient balancing of exploitation and exploration of the environment.   In ADAPT Phase II, the ADAPT agents will be enhanced to support explicit bi-directional interaction between the active inference model and human team members, which is essential for both increasing the application of the model in real-world settings. Aptima will extend the active inference model developed on the DARPA A-Teams program and the ADAPT Phase I, develop methods for communication between humans and agents, and validate of the combination of model and interaction methods in a Minecraft-based urban search-and-rescue synthetic task environment that is being jointly developed by Aptima and Professor Nancy Cooke’s lab at ASU for the DARPA ASIST program.