Mechanical Engineering

Accelerating Stochastic Simulation and Control with Learned Models of Mean Behavior

November 15, 2019

1:00 p.m. - 2:00 p.m.

160 Hall of Arts

We will show how learned models and efficient sampling algorithms can be used to accelerate the numerical estimation of the mean behavior of stochastic systems, without introducing any approximation error. Two expensive stochastic simulations will be considered: atmospheric aerosols simulated directly at the particle level, and reinforcement learning for legged robot control. The aerosol simulations follow billions of particles in a 3D region, such as the Northern California DOE CARES campaign area, to enable the direct computation of climate-impacting aerosol optical and cloud condensation nuclei properties. For learning robot control, we simulate jumping and landing dynamics on the microsecond timescale to train nonlinear neural network feedback policies for single-leg jumping.

Speaker: Matthew West, Associate Professor, Department of Mechanical Science & Engineering, University of Illinois at Urbana-Champaign