Direct learning neural and fuzzy control of two unknown dynamic systems using Powell optimization and genetic algorithms

This study establishes the utility of Powell's algorithm and Genetic algorithms for direct learning neural network and fuzzy logic controllers. The salient characteristics of these controllers offer: (1) learning capabilities, to improve plant tracking performance from one repetition to the next, (2) that required little or no prior knowledge about plant dynamics, (3) with no need for on-line plant identification, (4) as well as automatic controller parameter and structure learning.;The aforementioned controllers are validated with two non-linear plants. These plants are a simulated double pendulum and the x-joint of an IBM Eaglet II robot. Additionally, an inverted pendulum is also used to validate the learning fuzzy controller as well.;The experimental and simulation results show that: (1) The utility of both Powell's algorithm and Genetic algorithms have enabled direct learning for a neural network controller. A neural network controller based on Powell's algorithm converges faster than the one based on Genetic algorithms, while the latter is capable of escaping a local minima. (2) Structure learning in a neural network controller can be accomplished with a two step approach that alternates between adding a neuron and controller parameter tuning. (3) The utility of Powell's algorithm has enabled direct tuning of a fuzzy logic controller's Control State Action table. (4) The utility of Powell's algorithm has enabled tuning of scaling factors, which are part of the fuzzy controller's parameters. (5) A non-linear autoregressive with exogenous input model identified by the Simple Genetic algorithm is adequate for modeling the dynamics of the x-joint of the Eaglet II robot.