Evolving prototype rules and genetic algorithm in a combustion control

All approaches to rule-based control have relied upon a pre-classification of the state-space for their success. Given a classification of the state-space, either an individual control action or a whole set of control actions can be learnt. But quality of control is limited by the quality of the state-space classification. When no classification is specified, a more challenging problem is how to arrive at the most suitable partitioning of the state-space with its associated control actions. The approach presented in this paper is to generate prototype rules. Instead of learning a control action for every point encountered, a genetic algorithm is used to learn control actions for a set of limited number of prototype states and the nearest neighbour matching is employed to decide which of the rules to fire in any particular situation when controlling a system. The example of a simulated multiple burner combustion optimization is used to demonstrate this approach