Design of heuristic fuzzy controller

We propose a method for obtaining fuzzy rules of a dynamic control system without using any plant model or control surface. We maintain a table of previous state-control pairs ⟨s(t_j)⟩, y(t _j)⟩, together with a fuzzy evaluation e(t_j) of the goodness of the control action y(t_j) at the state s(t _j). To select the control action y(t) at the current state s(t), we first find several state-control pairs which match well with s(t) and have good evaluations. Then, we compute y(t) by a heuristic adaptation of the control actions of those neighboring points. When several successful control runs with different start-states have been obtained, we apply fuzzy clustering to all the points ⟨s(t_j ), y(t_j)⟩ accumulated so far to obtain the final set of rules, one from each cluster. These rules provide a good generalization of the points ⟨s(t_j), y(t_j)⟩, and they can be used for successfully controlling the plant for new start states. The effectiveness of our method is demonstrated by applying it to the inverted pendulum problem