摘要:Artificial intelligence systems,machine learning improves system performance automatically from experience.provides tools for intelligent data analysis.learning task category:supervised learning、partially supervised learning、reinforcement learning、unsupervised learning.a typical formulation:value function approximation directly find the parameters of the policy.supervised/partially supervised learning.machine learning=representation + evaluation + optimization.highly complex functions vs. gradient-based methods,relax to convex problems,convex loss functions are noise-sensitive [Long and Servedio, MLJ'OO],convex regularizations are not consistent [Fan and Li, JASA'O1],hard to be relaxed.