Doubly Regularized Maximum Margin Planning Algorithms Research

Maximum Margin Planning(MMP) model can convert the planning problem of Aitificial Intelligence(AI) to structured predcition problem in machine learning. The main idea of algorithm is to model the human being’s teaching behaviors, so that we can more easily impart our experience to machine. However, traditional MMP algorithms will be influenced by many irrelevant features when appling to the problems of high dimension feature. In this paper, we combine the linear MMP with feature selection problem and propose doubly regularized MMP algorithm to solve those highly dimensional problems.This paper surveys the MMP and feature seletion problem and then introduces the background knowledge of MMP and feature selection problem. From the objective func-tion of MMP, we can view this problem as a regularization of the structure risk function minimization problem. In order to automatically remove irrelevant feature in the learning process, we add a penalty term which can induce feature selection in the objection function and construct a double regularization of the structure risk function minimazation problem. According to the diffirent sparse penalty term we added, this paper proposes two novelty doubly regularized MMP model. One is based on lasso and the other one is based on MCP. We use subgradient descent method and augmented lagrangian multiplier method to solve those two models, and inference the algorithm 5 and algorithm 6 respectively. Theoretically, we proved the convergence of two algorithms mentioned above. Finally, we developed a test experiment system to validate the proposed algorithms.