โ€œMachine Learningโ€์„ ๊ณต๋ถ€ํ•˜๋ฉด์„œ ๊ฐœ์ธ์ ์ธ ์šฉ๋„๋กœ ์ •๋ฆฌํ•œ ํฌ์ŠคํŠธ์ž…๋‹ˆ๋‹ค. ์ง€์ ์€ ์–ธ์ œ๋‚˜ ํ™˜์˜์ž…๋‹ˆ๋‹ค :)

3 minute read

โ€œMachine Learningโ€์„ ๊ณต๋ถ€ํ•˜๋ฉด์„œ ๊ฐœ์ธ์ ์ธ ์šฉ๋„๋กœ ์ •๋ฆฌํ•œ ํฌ์ŠคํŠธ์ž…๋‹ˆ๋‹ค. ์ง€์ ์€ ์–ธ์ œ๋‚˜ ํ™˜์˜์ž…๋‹ˆ๋‹ค :)

Introduction to Markov Property

Definition. Markov Property

A random process has the <Markov property> if the conditional probability distribution of future states of the process depends only upon the present state.

For sequential states $S_i$, the conditional probability follows

\[p(S_{t+1} \mid S_0, S_1, \dots, S_t) = p(S_{t+1} \mid S_t)\]

์ด๋Ÿฐ <Markov property>๋ฅผ ๊ฐ€์ง„ Random Process๋ฅผ <Markov Process>๋ผ๊ณ  ํ•œ๋‹ค! <Markov Chain; ๋งˆ๋ฅด์ฝ”ํ”„ ์—ฐ์‡„>, <Brownian motion> ๋“ฑ์ด <Markov Process>์— ์†ํ•œ๋‹ค.


vs. Memoryless Property

<Markov Property>๋Š” ๊ณผ๊ฑฐ(past)์˜ ์ƒํƒœ์™€ ๋…๋ฆฝ์ด๋ผ๋Š” ์„ฑ์งˆ ๋•Œ๋ฌธ์— โ€œmemorylessโ€๋ผ๊ณ  ํ•œ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ, โ€œํ™•๋ฅ ๊ณผ ํ†ต๊ณ„(MATH230)โ€ ์ •๊ทœ ์ˆ˜์—… ์‹œ๊ฐ„์—๋Š” <Memoryless property>๋ฅผ ์•„๋ž˜์™€ ๊ฐ™์ด ์ •์˜ํ–ˆ๋‹ค.

\[P(X > n + m \mid X \ge m) = P(X > n)\]

๋‘ ๊ฐœ๋…์€ ์„œ๋กœ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์ผ๊นŒ? ์ผ๋‹จ ๋‘ ๊ฐœ๋…์ด ์ •์˜๋œ ์ƒํ™ฉ์„ ๋ช…ํ™•ํžˆ ๋ด์•ผ ํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค.

<Markov Property>๋Š” Random Process, ์ฆ‰ sequence of RV์˜ dependency์— ๋Œ€ํ•ด ๋ฌ˜์‚ฌํ•˜๊ณ  ์žˆ๋‹ค. ๋ฐ˜๋ฉด์— <Memoryless Property>๋Š” ์–ด๋–ค RV์˜ ๋ถ„ํฌ์— ๋Œ€ํ•ด ๋ฌ˜์‚ฌํ•˜๊ณ  ์žˆ๋‹ค. ์ฆ‰, ๋‘˜์ด ์ •์˜๋œ ์ƒํ™ฉ ์ž์ฒด๊ฐ€ ๋‹ค๋ฅด๋‹ค.

์ฆ‰, <Markov Property>๋Š” Random Process๊ฐ€ ๊ฐ–๋Š” Memoryless์— ๋Œ€ํ•œ ์„ฑ์งˆ์ด๋‹ค. <Memoryless Property>๋Š” Random Value๊ฐ€ ๊ฐ–๋Š” Memoryless์— ๋Œ€ํ•œ ์„ฑ์งˆ์ด๋‹ค.


Markov Chain

<Markov Chain; ๋งˆ๋ฅด์ฝ”ํ”„ ์ฒด์ธ, ๋งˆ๋ฅด์ฝ”ํ”„ ์—ฐ์‡„>๋Š” <Markov Property>๋ฅผ ๊ฐ€์ง„ โ€˜Discreteโ€™ Random Process๋ฅผ ์˜๋ฏธํ•œ๋‹ค.

๋งŒ์•ฝ ์ผ๊ธฐ์˜ˆ๋ณด๋ฅผ <Markov Chain>์œผ๋กœ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด, ์˜ค๋Š˜์˜ ๋‚ ์”จ๋ฅผ ํ†ตํ•ด ๋‚ด์ผ์˜ ๋‚ ์”จ๋ฅผ ํ™•๋ฅ ์ ์œผ๋กœ ์˜ˆ์ธกํ•˜๊ณ , ๋‹ค์‹œ ๋‚ด์ผ์˜ ๋‚ ์”จ ์ •๋ณด๋ฅผ ํ†ตํ•ด ๋ชจ๋ ˆ์˜ ๋‚ ์”จ๋ฅผ ํ™•๋ฅ ์ ์œผ๋กœ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋œ๋‹ค.

<Markov Process>์—์„œ๋Š” ์ƒํƒœ(State) $S$๋ฅผ ํ†ตํ•ด ๊ฐ ๊ณผ์ •์„ ๋ฌ˜์‚ฌํ•˜๊ณ , ๋˜ ๊ฐ ์ƒํƒœ์—์„œ์˜ ์ „์ด(transition)์— ๋Œ€ํ•œ ํ™•๋ฅ ์„ ์•„๋ž˜์™€ ๊ฐ™์ด <transition diagram>์œผ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋‹ค.

๋˜๋Š” transition์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ํ–‰๋ ฌ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜๋„ ์žˆ๋‹ค. ์ด๋ฅผ <transition matrix; ์ „์ด ํ–‰๋ ฌ>์ด๋ผ๊ณ  ํ•œ๋‹ค.

์˜ค๋Š˜/๋‚ด์ผ ๋ง‘์Œ ํ๋ฆผ
๋ง‘์Œ 0.3 0.7
ํ๋ฆผ 0.4 0.6

<Markov Chain>์˜ ์žฅ์ ์€ <transition matrix>๋ฅผ $n$๋ฒˆ ๊ณฑํ•ด $n$์ผ ํ›„์˜ ์ƒํƒœ๋ฅผ ์ „์ด ํ™•๋ฅ ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค! ์˜ˆ๋ฅผ ๋“ค์–ด ๋ชจ๋ ˆ์˜ ์ „์ด ํ™•๋ฅ ์„ ์•Œ๊ณ  ์‹ถ๋‹ค๋ฉด, <transition matrix> $P$๋ฅผ ๋‘๋ฒˆ ๊ณฑํ•˜๋ฉด ๋œ๋‹ค.

\[P^2 = P P = \begin{pmatrix} 0.3 & 0.7 \\ 0.4 & 0.6 \end{pmatrix} \begin{pmatrix} 0.3 & 0.7 \\ 0.4 & 0.6 \end{pmatrix} = \begin{pmatrix} 0.37 & 0.63 \\ 0.36 & 0.64 \end{pmatrix}\]

์—ฌ๊ธฐ๊นŒ์ง€์˜ ์„ค๋ช…์€ <Markov Chain>์— ๋Œ€ํ•œ ์ •๋ง ์‰ฝ๊ณ  ๊ฐ„๋‹จํ•œ ์„ค๋ช…์ด๋‹ค. ์œ„์˜ ์˜ˆ์‹œ๋Š” State $S$๊ฐ€ Discrete์ด๊ณ , Time ์—ญ์‹œ Discrete์ด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ <Markov Chain>์˜ ๊ฐœ๋…์„ ํ™•์žฅํ•˜๋ฉด, State-space๊ฐ€ continuous์ด๊ฑฐ๋‚˜ Time-space continuousํ•œ ๊ฒฝ์šฐ๋„ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ๋‹ค. ๐Ÿคฉ ๋” ์ž์„ธํ•œ ๋‚ด์šฉ์€ Wikipedia - Markov Chain์„ ์‚ดํŽด๋ณด์ž.


์š”์•ฝ.

  • <Markov Property>: Random Process์—์„œ์˜ Memorylessness์— ๋Œ€ํ•œ ์„ฑ์งˆ
  • <Markov Chain>: Discrete Random Process with Markov Property
  • <Markov Model>: <Markov Chain>์„ ๋”ฐ๋ฅด๋Š” ๋ชจ๋ธ

๋” ์ฝ์„๊ฑฐ๋ฆฌ.


References