An algorithm for goal-driven simulation

This paper addresses the problem of goal-driven simulation. Goal-driven simulation is a task frequently performed by users of simulation systems. It consists in determining, when possible, an assignment of one or several decision variable(s) in order to obtain a particular value for a specific goal variable. This task is poorly supported in simulation systems because of lack of appropriate algorithms. Some systems assist goal-driven simulation with a functionality called target value computation. This functionality allows users to set a value for a goal variable and to get the value of a decision variable by running a simulation “backwards” from this goal. However, target value computation is insufficient in current simulation systems: it does not deal with models involving conditional expressions in equations-a common case in practice-nor with under- and over-constrained problems, which frequently occur during goal-driven simulation. We present an algorithm which overcomes these difficulties. We propose to combine graph theoretic methods for monitoring the numerical solving process of the model and interval constraint reasoning for dealing with under-constrained and over-constrained problems. This algorithm, implemented in a simulation environment called AMIA, has been successfully applied to several large models containing thousands of equations