Joint constraint modelling using evolved topology generalized multi-layer perceptron (GMLP)

The accurate simulation of anatomical joint models is important for both medical diagnosis and realistic animation applications. Quaternion algebra has been increasingly applied to model rotations providing a compact representation while avoiding singularities. This paper describes the application of artificial neural networks topologically evolved using genetic algorithms to model joint constraints directly in quaternion space. These networks are trained (using resilient back propagation) to model discontinuous vector fields that act as corrective functions ensuring invalid joint configurations are accurately corrected. The results show that complex quaternion-based joint constraints can be learned without resorting to reduced coordinate models or iterative techniques used in other quaternion based joint constraint approaches.