[ML] 머신러닝의 학습 방법 (Supervised Learning, Unsupervised Learning, Reinforcement Learning)

머신러닝의 학습 방법

1. Supervised Learning (지도 학습)

데이터에 대한 Label(명시적인 정답)이 주어진 상태에서 컴퓨터를 학습시키는 방법이다. 훈련 데이터(Training Data)로부터 하나의 함수를 유추해내기 위한 방법으로, 이렇게 유추된 함수 중 연속적인 값을 출력하는 것을 회귀분석(Regression)이라 하고 주어진 입력 벡터가 어떤 종류의 값인지 표식하는 것을 분류(Classification)라 한다.

 

• Supervised learning decodes known relationships between independent variables and the dependent variable.
• This involves feeding the machine sample data with various features (X) and their known output value (y).
• The fact that the input and output values are known qualifies the dataset as “labeled” or “supervised.”

2. Unsupervised Learning (비지도 학습)

데이터에 대한 Label(명시적인 정답)이 주어지지 상태에서 컴퓨터를 학습시키는 방법으로, 데이터가 어떻게 구성되었는지를 알아내는 문제의 범주에 속한다. 이 방법은 지도 학습(Supervised Learning) 혹은 강화 학습(Reinforcement Learning)과는 달리 입력값에 대한 목표치가 주어지지 않는다. 

 

• The dependent variables aren’t known or labeled and the model looks at patterns among independents variables to create a new output.
• In the case of clustering analysis, this can be achieved by grouping similar data points and finding connections that generalize patterns, such as the grouping of suburbs with two-bedroom apartments that generate a high property valuation.
• As there are no known output observations available to check and validate the model, there is no true output in unsupervised learning and predictions are more subjective than that of supervised learning.
• Unsupervised learning is useful in situations where there’s no single clear prediction goal and exploratory data analysis is required to uncover new categories and subgroups.
• Other popular unsupervised learning tasks include anomaly detection such as fraudulent transactions or catching manufacturing defects, and automatically removing outliers and complexity from a dataset before feeding the data to a supervised learning algorithm.

3. Reinforcement Learning (강화 학습)

행동심리학에서 영감을 받았으며, 어떤 환경 안에서 정의된 에이전트가 현재의 상태를 인식하여, 선택 가능한 행동들 중 보상을 최대화하는 행동 혹은 행동 순서를 선택하는 방법이다. 

 

• The third and most advanced category of machine learning and is generally used for performing a sequence of decisions, such as “playing chess” or “driving an automobile”.
• Reinforcement learning is the opposite of unsupervised learning as the output (y) is known but the inputs (X) are unknown.
• The output can be considered as the intended goal (i.e. win a game of chess)
• the optimal input is found using a brute force technique based on trial and error.
• In the case of self-driving vehicles, movements to avoid a crash are graded positively. in the case of chess, moves to avoid defeat are rewarded.
• Over time, the model leverages this feedback to progressively improve its choice of input variables to achieve its desired output goal.