We are proposing a cognitive agent architecture that leverages Interactive Machine Learning, which will allow ground controllers and crewmembers the ability to train and retrain their cognitive agents during a mission based on new/novel mission data. Interactive Machine Learning (IML) is a human-centered paradigm in which end-users, e.g. crewmembers and ground-based Subject Matter Experts, iteratively build and refine the ML model through iterative cycles of input and review. Model refinement is driven by user input that may come in many forms, such as onboard data and communications, crew preferences, modified mission parameters along with a description of features and selection of high-level model parameters. IML is distinct from classical machine learning in that human intelligence is applied through iterative teaching and model refinement in a relatively tight loop. With this approach, our goal is to develop an architecture that provides end-users with the ability to interactively explore and adapt the training space with the goal of guiding the adaptation of the cognitive agent toward an intended behavior. This approach will allow crewmembers the ability to control how the cognitive agent learns and adapts during a long duration exploration mission, assuring that its performance improves and does not degrade over time. This work will present crewmembers and SME centered approach to applying IML methods to the design of a system that learns and adapts based on crewmembers inputs to the system. By leveraging IML we are able to address one of the primary challenges associated with the use of cognitive agents during a long-duration mission, the inability to adapt and learn from observation, instruction and interaction as missions proceed.
We expect the IML approach to developing training models for multi-agent assignment along with the ability for end-users to retrain the system will be of interest to a number of groups within NASA, e.g., Gateway habitat. Of course, future Mars expeditions could certainly make use of our cognitive agent, as those are the types of scenarios we’ve used in our development. In particular, our agent will be of interest at JSC to the EVA Exploration Office, the EVA Strategic Planning and Architecture group, and the Exploration Mission Planning Office.
The proposed cognitive architecture will benefit a number of TRACLabs commercial customers. For example, Baker Hughes has already expressed interest in licensing some of the new capabilities being developed in previous cognitive agent efforts, particularly the ontology and anomaly management aspects. We expect the ability for end-users to direct the adaptation of the system will be of interest.