MIXTAPE: Middleware for Interactive XAI with Tree-based AI Performance Evaluation

Kitware, in partnership with Penn State University, is proud to propose this Phase I STTR effort to create XAI middleware that can support the interactive explanation and visualization of AI decision-making systems. Current (X)AI algorithms are often application-specific, limiting their portability to cross-cutting domains. Our approach, MIXTAPE, will create modular and extensible software interfaces and implementations of different XAI tools for humans to evaluate the performance of their AI agents in specific environments (e.g., in Multi-Domain Operations). This work builds upon our previous experience on the DARPA XAI and XAI Toolkit efforts, where we developed novel techniques for explaining the complex reasoning of AI agents in different environments. For explanations, we will extend the tree-based visualizations we used previously, which can show current and predicted future action and state information at critical decision points. To provide a process for humans to scaffold their explanation consumption, we will extend our work on After-Action Review for AI. Given the flexible nature of our framework, additional forms of explanation can also be incorporated, such as saliency maps, tracking win/loss probabilities over time, and including geospatial visualization when appropriate. We will first test our approach on the ARL Battlespace environment, but it can apply to different AI agents, wargaming scenarios, and AI testbeds.