Integrating Episodic and Semantic Memory in Machine Teammates to Enable Explainable After-Action Review and Intervention Planning in HAA Operations

Abstract

A critical step to ensure that AI systems can function as effective teammates is to develop new modeling approaches for AI based on the full range of human memory processes and systems evidenced by cognitive sciences research. In this paper we introduce novel techniques that integrate episodic and semantic memory within Artificially Intelligent (AI) teammates. We draw inspiration from evidence that points to the key role of episodic memory in representing event-specific knowledge to enable simulation of future experiences, and evidence for a representational organization of conceptual semantic knowledge via self-organizing maps (SOMs). Together, we demonstrate that these two types of memory working in concert can improve machine capabilities in co-learning and co-training scenarios. We evaluate our system in the context of simulated helicopter air ambulance (HAA) trajectories and a formal model of performance and skill, with interventions to enable an AI teammate to improve its capabilities on joint HAA missions. Our modeling approach contrasts with traditional neural network training, in which specific training data is not preserved in the final trained model embedding. In contrast, the training data for our model consists of episodes containing spatial and temporal information that are preserved in the model’s embedding. The trained model creates a structure of relationships among key parameters of these episodes, allowing us to understand the similarity and differences between performers (both human and machine) in outcomes, performance, and trajectory. We further extend these capabilities by enhancing our semantic memory model to encode not just a series of episodes, but labeled directed edges between regions of semantic memory representing meta-episodes. These directed edges represent interventions applied by the performer to improve future episodic outcomes in response to identified gaps in capability. These interventions represent the application of specific co-training strategies as a labeled transition system, linking episodes representing pre-intervention and post-intervention performance. This allows us to represent the expected impact of interventions, simulating improvements and skill decay, providing the machine with team-aligned goals for self-improvement between episodes to positively impact future teamwork.

Keywords: Human-Machine Teaming, Co-Learning, Co-Training, After-Action Review, Cognitive Modeling, Ai, Episodic Memory, Semantic Memory

DOI: 10.54941/ahfe1005003