How blind can a blind fuzzy logic controller design be? Analysis of cell state space based incremental best estimate directed search algorithm

Reports a discovery that has never been reported before on the global performance of Tskagi-Sugeno (TS) type fuzzy logic controller (FLC) with random rule output function parameters. We found that under certain cell resolutions, the chance that the performance of a random controller is comparable with that of a man made controller is very high. A cell state space based TS type FLC automatic parameter optimization algorithm named incremental best estimate directed search (IBEDS) is presented with a detailed analysis of its capability of designing a FLC totally without any expert knowledge. Initially, IBEDS was suggested to start with an initial training set sampled from the control surface of a FLC with poor performance. Then another random FLC is trained in an iterative procedure by a least mean square (LMS) algorithm. In each iteration, the global and local performance of the trained FLC is evaluated, the training set is then updated based on the evaluation. The initial value of the training set may need expert knowledge. However, with the discovery on random FLCs, IBEDS is found to be able to bootstrap with an empty training set, and still reach an optimal solution within a reasonable number of iterations. The mechanism behind this phenomenon is analyzed in detail with 2D and 4D inverted pendulums. It is shown that with IBEDS, a totally blind FLC design without any expert knowledge can be done efficiently.