This document introduces deep reinforcement learning and provides some examples of its applications. It begins with backgrounds on the history of deep learning and reinforcement learning. It then explains the concepts of reinforcement learning, deep learning, and deep reinforcement learning. Some example applications are controlling building sway, optimizing smart grids, and autonomous vehicles. The document also discusses using deep reinforcement learning for robot control and how understanding the principles can help in problem setting.
This document introduces deep reinforcement learning and provides some examples of its applications. It begins with backgrounds on the history of deep learning and reinforcement learning. It then explains the concepts of reinforcement learning, deep learning, and deep reinforcement learning. Some example applications are controlling building sway, optimizing smart grids, and autonomous vehicles. The document also discusses using deep reinforcement learning for robot control and how understanding the principles can help in problem setting.
2. 1. SHAP
- 아이디어
한 obs에서 특정 Feature에 따라 Prediction 변화값을 여러 번 구하고 평균값 계산하자
1) i번째 관측치에서 feature_j의 효과를 구하자
- 예측모델 f에 feature_j 값에 따른 변화량 기록
)(ˆ)(ˆˆ 1111
ijiiij xxfxf
k
ijkij w ˆˆ
2) 1)을 K번 반복 후 가중평균을 계산
-> i번째 관측치에서 feature_j의 효과 (=SHAP value)
= SHAP value
3. 1. SHAP
- Cat-forbidden(feature_j) 이 주택가격예측(Prediction)에 주는 효과?
)(ˆ)(ˆˆ 1111
ijiiij xxfxf
- 너무 많은 Coalition Case를 어떻게 estimation 해야 할까?
-> Monte-Carlo, Kernel 등 방법 적용 중
]|)([)(ˆ
sx xxfESf
- Coalition S = a subset of the features
4. 1. SHAP
- Monte-Carlo method
ex. 원의 넓이를 구할 때, 사각형의 boundary를 정해준 뒤 random sampling
원 안에 점이 포함된 비율 * 사각형의 넓이 = 원의 넓이
5. 1. SHAP
- 알고리즘 (Monte-Carlo method)
1) Require
X = Data(n x p) / x = obs / f = ML model / K = 반복 수
2) Iterations
},,1{ pj
},,1{ Kk
For all
For all
- Data 에서 z번째 instance random sampling
- p개의 feature를 나열하는 방법(순열) 중 하나 randomly select = ordering method
- x와 z를 ordering
- Construct two new instances
),,,,,,( 111
*
pjjj
j
zzxxxx
),,,,,,( 111
*
pjjj
j
zzzxxx
),,,,( 1 pjo xxxx ),,,,( 1 pjo zzzz
)(ˆ)(ˆˆ **)( jjk
ij xfxf
- Calculate difference
-> ordering을 이용해 coalition 정의
6. 2. Tree SHAP
- Ensemble Tree model의 결과를 더 잘 설명하기위한 Tree SHAP algorithms
1) Leaf-node 이면
w*v 값 반환
2) internal-node 일 때 split feature가
- coalition S 에 포함되면 threshold기준으로
좌우 중 한가지 노드 선택
- 그렇지 않으면 좌우 노드 모두 계산
단 가중치가 감소
f1
f2
f3
< 3
< 6
< 1
Coalition S = {f1, f2}
x = {2,4,6}
Score
0.9
0.1 0.3
0.2
Tree 1
1
ˆij 값 계산
7. 2. Tree SHAP
- Weight update
m = the path of unique features we have split on so far
m의 가중치를 계속 업데이트 (EXTEND , UNWIND)
- Ensemble Tree model의 결과를 더 잘 설명하기위한 Tree SHAP algorithms