Intelligent control system for CFD modelling software
In this thesis we show that it is possible to create an intelligent agent capable of emulating the human ability to control CFD simulations and provide similar benefits in terms of performance, overall reliability and result accuracy. We initially consider the rule-based approach proposed by other researchers. It is argued that heuristic search is better suited to model the techniques used by human experts. The residual graphs are identified as the most important source of heuristic information relevant to the control decisions. Three different graph features are found to be most important and dedicated algorithms are developed for their extraction. A heuristic evaluation function employing the new extraction algorithms is proposed and implemented in the first version of the heuristic control system (ICS 1.0). The analysis of the test results gives rise to the next version of the system (ICS 2.0). ICS 2.0 employs an additional expert system responsible for dynamic pruning of the search space using the rules obtained by statistical analysis of the initial results. Other features include dedicated goal-driven search plans that help reduce the search space even further. The simulation results and overall improvements are compared with non-controlled runs. We present a detailed analysis of a fire case solution obtained with different control techniques. The effect of the automatic control on the accuracy of the results is explained and discussed. Finally, we provide some indications for further research that promise to provide even greater performance gains.