Forward Models for Non-Convex Data

 

A forward model is a learned approximation of the physical environment and it is used in series with an inverse model to form a composite learning system. This learning system is capable of solving the inverse mapping for non-convex regions of the solution space,[, ]. The training technique for the composite model is called distal learning and it is illustrated in Figure . The learner controls a distal outcome via a set of proximal variables which are inputs to a physical environment, in our case a physical model of two violin strings. The variable names and their functions for the composite system are outlined in Table .\

  
Figure: Distal Learning

  
Table: Simulation Input and Output Variables

• Training the Forward Model
• Training the Inverse Model using the Forward Model