Causal Adaptive Decision Aid (CADA)
Abstract
Logistics and planning personnel are overloaded by the increasing number of high dimensional data layers they are expected to analyze in short amounts of time. The goal of the Causal Adaptive Decision Aid (CADA) is to facilitate the foraging and sensemaking steps of their analyses by rapidly summarizing these layers and displaying the cause-and-effect concepts which are relevant to recommend explainable courses of action (COAs). A causal model for COA recommendation has the benefits of being more robust. The goal of this study is to simulate realistic naval scenarios, from noisy sensor data to human decision, to be used as a benchmark data set for subsequent tasks. We detailed our approach for the various software modules that comprise the CADA system. It should be noted that the existence of cause-and-effect concepts in any system of high-dimensional variables is not guaranteed. However, we have shown that in datasets that do contain cause-and-effect concepts, it is possible to utilize some amount of human supervision to enhance the correlation-based representations of discovered concepts such that they more concisely represent the known causal dynamics of the system. Among the lessons learned on this project, we have found that one of most challenging aspects of causal inference and modeling research is finding real-world observational datasets with known causal ground truth. This is a major limitation in performing the type of thorough testing and evaluation that would enable transitioning cutting-edge causal theory to applications of naval relevance. Despite this issue, we are optimistic about the future of applied causal inference due recent calls for benchmark datasets and real-world applications.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 19, 2023
- Accession Number
- AD1206088
Entities
People
- Aruna Jammalamadaka
- Rajan Bhattacharyya