Prediction of Physics Simulation using Dimensionality Reduction and Regression
Robert Dupre, Vasileios Argyriou, Darrel R. Greenhill · Journal of Graphics Tools · 2013
In this article two novel approaches for physics simulation prediction are proposed with applications in graphics rendering and multimedia, aiming to reduce the computational cost of simulating physics in 3D environments. Firstly a novel use of machine learning techniques is combined with an innovative regression based prediction mechanism. Secondly dimensionality reduction is employed to redefine simulation data into a linear representation which is subsequently used in a classification method. Finally a prediction based framework is proposed utilising these methods which, based on trained models, produces visually similar effects without the need for frame by frame simulation. Extensive experimentation is carried out in both simple and complex simulation scenarios to evaluate the performance of the proposed prediction system. Results are presented indicating that in cases where precision is not essential, our system can be utilized providing visually similar kinematics whilst reducing computation time.