PhD Dissertation Proposal Defense: Pracheta Amaranath, The Interface of Simulation and Causal Modeling
Content
Speaker
Abstract
This thesis investigates the interplay between simulation and causal inference, focusing on how causal modeling can enhance simulation and vice versa. The first part develops modular dynamic Bayesian networks (MDBNs) as causal metamodels for discrete-event simulations of Markovian queueing networks. MDBNs approximate the temporal evolution of the system states, enabling accurate and efficient inference of probabilistic and causal queries. Building on this foundation, the thesis explores scalable inference strategies for MDBNs, including recurrence equations that approximate state transitions in the system, approximate inference algorithms, and implementations leveraging modern computing techniques to extend applicability to more complex systems.
The second part examines the inverse problem: using simulation models as generative methods that produce synthetic datasets reflecting the underlying causal relationships of a real-world data-generating process. An empirical evaluation of existing generative neural network methods reveals that small sample sizes, architectural biases, and fixed parameter specifications can produce inconsistent datasets and unreliable estimator evaluations. To address these issues, the thesis introduces simulation-based inference for causal evaluation (SBICE), a Bayesian framework that models simulator parameters as uncertain, infers a posterior distribution over configurations compatible with the observed data, and generates synthetic datasets grounded in both prior knowledge and empirical evidence. This approach supports principled sensitivity analyses and improves the robustness of causal estimator benchmarking.
Advisors
Peter Haas and David Jensen