【Machine Learning】监督学习 非监督学习及强化学习对比

Supervised LearningUnsupervised LearningReinforced Learning

Goal：

How to apply these methodsHow to evaluate each methods

What is Machine Learning?

putational statistics

putational artifacts(人工制品) that learn over time based on experience

一、分类

Supervised LearningUnsupervised LearningReinforcement Learning

1.1 Supervised learning——Approximation

一句话实质：About Function Approximation(函数逼近)，or Approximate function induction(近似函数归纳)feed with labeled examples，comeing up with some function that generalizes beyond(泛化函数)有反馈

1.2 Unsupervised learning——Description

一句话实质：About Compact(简洁的) Description无监督学习是密切相关的统计数据密度估计的问题。无反馈Unsupervised learning could be helpful in the supervised Setting

1.3 Reinforcement learning (增强学习)

一句话实质：Learning from delayed reward (通过延迟性奖励进行学习)执行许多步之后才知道反馈，就像下棋（对比监督学习的立即反馈）

二、归纳法(induction)与演绎法(deduction)

Generalize 泛化了解机器学习发展史机器学习算法与归纳而不是演绎有关Inductive bias 归纳偏差

归纳：从示例到一般规律（从一个示例得出更普遍的规律）

演绎：从规则到实例，a general rule to specific instances,basically like reasoning(推理)

三、三种机器学习的比较

表述成：优化问题

Supervised Learning —— labels data well(to find a funtion to score that) （标记数据）

Unsupervised Learning —— cluster scores well(最好的分类方法)

Reinforcement learning —— behavior scores well (最好的表现)

3.2 Data

Data is king in machine learning.

转变：以算法为中心——》以数据为中心

Believe in your data！