When to Observe or Act? Interpretable and Causal Recommendations in Time-Sensitive Dilemmas

Decision-making is complicated when only uncertain or partial information relevant to the optimal choice is known and the ability to obtain this information is constrained by finite resources like time. Yet, being sensitive to implicit information and competing priorities related to a decision may demand delayed choice for the sake of obtaining vital information, e.g., firefighters risking the spread of a fire by delaying countermeasures before determining if the building is free of civilians. This work seeks to provide a framework for recommendations that can address these challenges in real-time scenarios represented by a sequential decision problem in which agents may (1) perform investigations to obtain information about variables important to (2) interventions that are graded by a multi-dimensional utility function. We demonstrate one approach that navigates these scenarios using a Causal Decision Network (CDN) to distinguish interventional from non-interventional options available to the agent and the consequences to implicit utility and value of information that each provide. Simulations demonstrate the efficacy of this approach and contextualize this beginning framework in a larger context for future directions of real-time recommenders.