Learning automata is one of the important artificial intelligencealgorithms, it operates in the probability space, whether the sample isequilibrium or not will not affect the performance. Due to the advantagessuch as robustness to the noise, the global optimization ability, adaptivelearning property and so on, LA can perform well in various stochasticenvironments, which shows the importance for deep research.Firstly, by researching and summarizing the existing algorithms of thelearning automata, we present a novel algorithm named Stochastic EstimatorReward-Penalty Learning Automata (SERP). This algorithm is based on thealgorithm with Stochastic Estimator named SERI, which modifies the way ofupdating the action selection probability. The detail of SERPalgorithm hasbeen discussed and the convergence property of it has also been verifiedthrough experiment. Moreover, the application of learning automata in stochastic pointlocation is also investigated in this thesis. And the algorithm introducedabove has been used for solving this problem continuously. The experimentshows the great converging property of this method.In addition, a new algorithm named Adaptive Step Searching is proposedto solve the SPL problem by discretizing the searching region. The core pointis to adapt the step size during the searching to approach the optimal pointmore accurately and faster. The procedure and the experiments of the ASSmethod are shown in the thesis as well. |