๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

๐“ก๐“ธ๐“ธ๐“ถ5: ๐’ฆ๐‘œ๐“‡๐‘’๐’ถ ๐’ฐ๐“ƒ๐’พ๐“‹/Artificial Intelligence(COSE361)

[์ธ๊ณต์ง€๋Šฅ] 15. Probabilistic Reasoning over Time (PRoT)

1. Introduction

 

 - ์„ธ์ƒ์„ ๋ช‡ ๊ฐœ์˜ observable ํ•˜๊ฑฐ๋‚˜ ๊ทธ๋ ‡์ง€ ์•Š์€ random variables์„ ํฌํ•จํ•˜๊ณ  ์žˆ๋Š” ์–ด๋–ค snapshot ๋˜๋Š” time slices์˜ ์ง‘ํ•ฉ์ด๋ผ๊ณ  ์ƒ๊ฐํ•˜์ž.

 - Xt : state variables at time t , ์‹œ๊ฐ„์ด t์ผ ๋•Œ์˜ state variable๋“ค์˜ ์ง‘ํ•ฉ. ๊ด€์ธกํ•  ์ˆ˜ ์—†๋Š”, ๊ทธ ๊ฐ’์„ ์•Œ ์ˆ˜ ์—†๋Š” variable์ด๋‹ค.

 - Et : evidence variables, ๊ด€์ธกํ•  ์ˆ˜ ์žˆ๋Š”, ์•Œ ์ˆ˜ ์žˆ๋Š” variable

 - transition model : state์ด ์–ด๋–ป๊ฒŒ ๋ณ€ํ•˜๋Š”์ง€ ์ •์˜ํ•œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด t-1 ์ผ ๋•Œ state์ด X์ผ ๋•Œ, ๋‹ค์Œ state์ด X ์ผ ํ™•๋ฅ 

 - sensor model : evidence variable์ด ๊ฐ’์ด ์–ด๋–ป๊ฒŒ ๋˜์–ด์•ผ ํ•˜๋Š”์ง€ ์ •์˜ํ•œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด state๊ฐ€ X์ผ ๋•Œ Evidence E๊ฐ€ ๊ด€์ธก๋  ํ™•๋ฅ 

 

  1) Transition Model : P(Xt | X 0:t-1) (state Xt์˜ ํ™•๋ฅ ๋ถ„ํฌ ์ •์˜) 

   - First-order Markov process : ์˜ค์ง ์ง์ „ state์—๋งŒ dependence ํ•˜๋‹ค 

     P(Xt | X 0:t-1 ) = P(Xt | Xt-1)

   - Second- order Markov process: ์ง์ „ 2๊ฐœ์˜ state์—๋งŒ dependenceํ•˜๋‹ค 

     P(Xt | X 0:t-1 ) = P(Xt | Xt-1, Xt-2)

 

   - Stationary process : ์‹œ๊ฐ„์ด ๋ณ€ํ•ด๋„ ํ™•๋ฅ ์€ ๋ณ€ํ•˜์ง€ ์•Š๋Š”๋‹ค. ์ฆ‰, ๋ชจ๋“  t์— ๋Œ€ํ•ด์„œ P(Xt|Xt-1) ์˜ ๊ฐ’์€ ๋Š˜ ๊ฐ™๋‹ค

   - ๋ชจ๋“  ๊ฒƒ์˜ ์‹œ์ž‘์€ P(X0) ; time์ด 0์ผ ๋•Œ์˜ prior probability

 

  2) Sensor Models : P(Et | X0:t, E1:t-1) = P(Et | Xt) (evidence์˜ ํ™•๋ฅ ๋ถ„ํฌ ์ •์˜)

   -> Sensor Markov assumption : Et๋Š” Xt์—๋งŒ dependent ํ•˜๋‹ค๊ณ  ๊ฐ€์ •ํ•œ๋‹ค (ํ˜„์žฌ state์—๋งŒ ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค) 

  

  3) Complete Joint Distribution 

   P(X_0:t , E_1:t) = P(X0)P(X1|X0)P(E1|X1, X0)P(X2|X1, X0, E1)P(E2|X2,X1,X0,E1) ...

  = P(X0)P(X1|X0)P(E1|X1)P(X2|X1)P(E2|X2) ... 

  = P(X0)∏P(Xi|X_i-1)P(Ei|Xi) (transition model * sensor model)

 

  4) Inference in Temporal Models

   โ‘  Filtering(=state estimation)

       :  1~t์˜ evidence ๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ Xt๋ฅผ ๊ตฌํ•˜๋Š” ๊ฒƒ. P(Xt | e_1:t) 

   โ‘ก Prediction 

      : 1~t์˜ evidence๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ ๋ฏธ๋ž˜์˜ state, t์‹œ๊ฐ„ ์ดํ›„์˜ state X_t+k๋ฅผ ์ถ”์ธกํ•˜๋Š” ๊ฒƒ 

       P(X_t+k | e_1:k)

   โ‘ข Smoothing 

      : 1~t์˜ evidence๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, t ์ด์ „ ์‹œ์ ์˜ state  ๊ตฌํ•˜๊ธฐ. P(Xk | e_1:t)

 

   โ‘ฃ most likely sequence of states : observations์˜ sequence๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ์ด ๊ด€์ธก ๊ฒฐ๊ณผ๋กœ๋ถ€ํ„ฐ ์ƒ์„ฑ๋  ํ™•๋ฅ ์ด ๊ฐ€์žฅ ๋†’์€ sequence of states๋ฅผ ๊ตฌํ•œ๋‹ค. arg max P(x_1:t | e_1:t) 

 

  -> transition and sensor models์„ ๋ชจ๋ฅด๋”๋ผ๋„ observation์„ ํ†ตํ•ด ํ•™์Šต๋  ์ˆ˜ ์žˆ์Œ. (EM algorithm)

 

 

2. Filtering : ํ˜„์žฌ ์‹œ๊ฐ„๊นŒ์ง€์˜ evidences๋ฅผ ์•Œ ๋•Œ ํ˜„์žฌ state ์ถ”์ธก

 P(Xt | e_1:t) = α Forward(f_1:t-1, et) 

 f_1:0 = P(X0 | e_1:0) = P(X0) // prior probability 

1~0 ์€ ์—ญ์ˆœ์ด๋‹ˆ๊นŒ ๋ง์ด ์•ˆ๋จ! ๊ทธ๋ƒฅ ํ•ญ์ƒ ํ™•๋ฅ  1์ด๋ผ๊ณ  ์ƒ๊ฐํ•˜์…ˆ ใ…‡ใ…‡

 

-> state variables์ด ๋‹ค discreteํ•  ๋•Œ, time๊ณผ space๋Š” constantํ•˜๊ฒŒ ํ•„์š”(?)ํ•จ. ๊ทธ๋ฆฌ๊ณ  ์ด ์ƒ์ˆ˜๊ฐ’์€ state space์™€ question์˜ temporal model์˜ ์ข…๋ฅ˜์— ๋”ฐ๋ผ ๊ฒฐ์ •๋จ. ์ฆ‰, time/space complexity๋Š” constantํ•˜๋ฉฐ, ์ด constant ๊ฐ’์€ state์™€ time t์— dependentํ•˜๋‹ค. 

 

 

3. Evidence

  - Evidence Sequence Likelihood 

   : P(e_1:t) = ∑ P(xt, e_1:t) 

 

4. Prediction : t๊นŒ์ง€์˜ evidence๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ t+k์˜ ๋ฏธ๋ž˜์˜ state์„ ์˜ˆ์ƒํ•˜๋Š” ๊ฒƒ 

P(X_t+k | e_1:t) 

 

5. Smoothing 

 -Backward Message : state k ์ผ ๋•Œ ๋ฏธ๋ž˜์˜ evidence e_k+1:t ๊ตฌํ•˜๊ธฐ

  P(e_k+1:t | Xk)

 -Smoothing : 1~t์˜ evidence๋ฅผ ์•Œ ๋•Œ t>k ์ธ ์‹œ์ , ์ด์ „ ์‹œ์ ์˜ state ์•Œ๊ธฐ

 

 

 

6. Viterbi Algorithm : ๊ฐ€์žฅ ํ™•๋ฅ ์ด ํฐ state sequence ๊ตฌํ•˜๊ธฐ 

  1) Lattice