Session: 02-03-03: Optimization

Paper Number: 145790

145790 - Improving Road Safety With Human-in-the-Loop Bayesian Optimization Using Driver Vision Obstruction Simulations 

When performing vision obscuration studies using Digital Human Modeling (DHM), one can evaluate  the percentage of driver visibility. Often, the vehicle’s A-pillar, the forwardmost structural component next to the front windshield, obscures the vision of the driver, making it an ideal component to redesign.  By introducing cutouts or see-through geometry in the A-pillar, we can increase the driver’s visibility of possible targets, which reduces the chance of an accident.  However, there are constraints on the design of the cutouts in the A-pillar to ensure the design is geometrically achievable and manufacturable .  Practically, determining the feasibility of a particular A-pillar design cannot be attained purely through mathematical or geometric models; it will be more effective to include designers in the loop to help assess feasibility. Designers can efficiently assess geometric layouts and manufacturability using their expert knowledge. Additionally, designing the A-pillar’s cutouts, evaluating its feasibility, and simulating the visibility is time intensive. This makes Bayesian optimization, a method  for optimizing expensive functions with minimal samples, an ideal optimization method to handle this problem. This study examines how altering the shapes, dimensions, and number of A-pillar cutouts can influence driver visibility under a broad array of driving conditions under these constraints. We aim to optimize the overall visibility score for a driver    given this variety of A-pillar configurations while ensuring that the design is feasible.  This is done by combining previous research in modeling driver visibility with surrogates and using Bayesian optimization with designer feasibility feedback. Surrogate models, computational representations of complex functions, are used to represent the driver visibility from the DHM simulation and design feasibility as a function of variable design parameters. A Bayesian optimization algorithm combines these two surrogates to optimize the visibility while ensuring that the final design is feasible. It sequentially provides the designer a batch of design points that potentially improve the design and have a high probability of feasibility; then the designer can evaluate the candidate design and provide feedback on feasibility. The results from the DHM simulation and feasibility are fed back into the Bayesian optimization algorithm, at which point a new batch of design points is created. A novel aspect of this approach is the ability to send the designer multiple designs at once, which is extremely useful as that minimizes the number of interactions they have with the algorithm.   Our analysis encompasses the performance of an SUV car model, evaluating different A-pillar geometries, their feasibility, and the resulting visibility in a daily traffic scenario. Our preliminary results show that surrogate models can effectively represent driver visibility in these scenarios and that designers can interact with Bayesian optimization to guide the design process. This is being used to find the optimal geometry for A-pillar cutouts that maximize the driver’s visibility.  

Presenting Author: Cole Jetton Oregon State University

Presenting Author Biography: Cole Jetton is a Ph.D. student at Oregon State University. He focuses on integrating designer knowledge into the optimization process, Bayesian optimization, and using Agent-Based Models to simulate device fleets.

Authors:

M. Amin Firouzi Oregon State University
Cole Jetton Oregon State University
Vignesh Bhaskaran Oregon State University
Christopher Hoyle Oregon State University
H. Onan Demirel Oregonstate University