Creating Models to Control Simulations: A Generic Approach

When things go wrong with complex systems, the consequences can be disastrous. Computer simulations are often used for training operators to control these systems and as a way of safely investigating system behaviour. If models of human operators can be used to drive these simulations, this should help to highlight possible system design problems in a more costeffective manner than having to use real operators. We are developing a generic approach for creating cognitive models that can learn to control any simulation that has its user interface written using a particular toolkit. This approach is based on using a robust cognitive architecture and a robust interface development tool. We identify a set of general requirements that the models should meet. An initial model, which satisfies a number of these, controls a simulation of a simple air traffic control task. The model learns while performing the task interactively and by reflecting on task performance afterwards.