Dynamic Discrete Choice Estimation with Partially Observable States and Hidden Dynamics

点击翻译按钮获取中文摘要

This award will investigate rigorous methods for data-driven model identification and estimation for stochastic dynamic decision processes. Many dynamical systems are modeled via discrete states that evolve via stochastic transitions and control inputs. Rather than assume a model structure in advance, this project develops methods to "learn" the model structure, including preferences of the controller and model states, from available data. The project will develop these new methods in the context of technology-assisted human performance in vehicle operations. Humans drivers operating vehicles on a busy highway often engage in secondary tasks that demand cognitive effort (e.g. attending to a business-related phone call, listening to a podcast). The driver is aware of their engagement in the secondary task but may be only imperfectly aware of its effect on their own ability to drive safely. On-board autonomous technology with the ability to accurately estimate the driver’s safety level may improve performance through auditory warnings or partial automation (e.g., adaptive cruise control). However, existing estimation methodologies are built on the assumption that the driver has perfect state observability, whereas the driving task’s overall level of safety is only imperfectly observed. This project provides a novel estimation method to build a predictive model of the agent’s driving behavior by considering imperfectly observed driving states (e.g. safe and unsafe driving). The research benefits the society by enabling on-board automation system to monitor the driver's activities and potentially intervene to improve driving performance and safety.This award will develop new methodologies and algorithms for learning a model of dynamic decisions with hidden states. The research draws on and generalizes dynamic discrete choice models to consider memoryless partially observable states. This project will also examine estimation of (non-exponential) semi-Markov hidden models with state-dependent sojourn time distributions. This research will rigorously examine whether and/or under what conditions the models are identifiable and ascertain the implications for robustness of estimation results. The new methods will leverage experimental collection of observable data known to correlate with distraction (e.g., breathing rate, heart rate variability, eye-tracking) together with observed actions (e.g., maneuvers) to estimate a compact model of the agent's control policies and the dynamics of partially observable states (safe and unsafe driving). The model and estimation approach will be validated through a high-fidelity driving simulation study.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.