Optimisation strategies for modelling and simulation
Progress in computation techniques had been dramatically reducing the difference between modelling and simulation. Simulation as natural outcome of modelling is used both as a tool to predict the behaviour of natural or artificial systems, a tool to validate modelling and a tool to build and refine models - in particular identify models internal parameters. In this paper we will concentrate upon the latter, model building and identification, using modern optimisation techniques, through application examples.
The first example will be taken from synthetic image animation. Realistic image animation requires going a step further than conventional artists-driven kinematics and using masses and forces. Unfortunately this is hard to obtain with most real world objects. We show how it is possible to learn the model's internal physical parameters from actual trajectory examples, using Darwin-inspired evolutionary algorithms.
In a second example, taken from image analysis, we show how it is possible to automatically learn an algorithm - here, a contour tracking algorithm - using Genetic Programming, another version of evolutionary optimisation.
In the third example, we will demonstrate how it is possible, in order to solve some problems where the expression of solutions it too complex to be easily handled by a reasonably simple optimisation technique, to split the problem into simpler elements which can be efficiently evolved by an evolutionary optimisation algorithm - what is now called ``Parisian Evolution''. This will be illustrated through the ``Fly algorithm'', a real-time stereovision algorithm which skips conventional preliminary stages of image processing, now applied into mobile robotics and medical imaging.
The main question is now, to which degree is it possible to delegate to a computer a part of the physicist's role, which is to collect examples and build general laws from these examples?