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

5 minute read

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


์ด๋ฒˆ ํฌ์ŠคํŠธ๋Š” โ€œํ™•๋ฅ ๋ก (Probability Theory)โ€๊ณผ โ€œMachine Learningโ€์—์„œ ๋“ฑ์žฅํ•˜๋Š” โ€œProcessโ€ ๊ฐ€ ๋ถ™์€ ๋ชจ๋“  ๊ฐœ๋…์„ ๋„“์€ ์‹œ์•ผ๋กœ ์‚ดํŽด๋ณด๊ธฐ ์œ„ํ•ด ์ž‘์„ฑํ•œ ํฌ์ŠคํŠธ์ž…๋‹ˆ๋‹ค. ๋‹ค๋ฃจ๋Š” ์ฃผ์ œ๋Š” ์•„๋ž˜์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.


Introduction to Random Process

Definition. Random Process

A random process is a time-varying function, that assigns the outcome of a random experiment to each time instant.

๋Œ€์ถฉ, (time instant)์— random experiment์˜ ๊ฒฐ๊ณผ(outcome)์„ ๋งคํ•‘ํ•œ๋‹ค๋Š” ๋œป์ด๋‹ค.

๋˜๋Š” ์•„๋ž˜์™€ ๊ฐ™์ด ์ •์˜ํ•˜๊ธฐ๋„ ํ•˜๋Š”๋ฐ,

A (infinite) sequence of random variables $X_1, X_2, \dots, X_n, \dots$

์ฆ‰, RV์˜ infinite sequence๋ฅผ <random process>๋ผ๊ณ  ํ•œ๋‹ค. ์ฒซ๋ฒˆ์งธ ์ •์˜๋ณด๋‹ค๋Š” ๋‘๋ฒˆ์งธ ์ •์˜๊ฐ€ ์ข€๋” ์™€๋‹ฟ๋Š” ํŽธ์ด๋‹ค. ๐Ÿ‘

<random process>๋ฅผ ์ •์˜ํ•  ๋•Œ, RV $X_i$๊ฐ€ ๋“ฑ์žฅํ–ˆ์œผ๋‹ˆ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์•„๋ž˜์˜ ์„ฑ์งˆ๋“ค์„ ๊ฐ€์งˆ ๊ฒƒ์ด๋‹ค.

  • $E[X_i]$: mean of RV
  • $\text{Var}(X_i)$: variance of RV
  • $p_{X_i} (x_i)$: marginal probability distribution of RV

์šฐ๋ฆฌ๋Š” <random process> ์ž์ฒด์˜ ๋ถ„ํฌ๋ฅผ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜๋„ ์žˆ๋Š”๋ฐ, ์ด๊ฒƒ์€ ์•„๋ž˜์™€ ๊ฐ™์€ joint probability distribution์ด ๋œ๋‹ค.

\[p_{X_1, \dots, X_n, \dots} (x_1, \dots, x_n, \dots)\]

๋˜, ์šฐ๋ฆฌ๋Š” <random process>์˜ Sample Space $\Omega$์— ๋Œ€ํ•ด ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ๋‹ค.

<random process> ์ค‘ ํ•˜๋‚˜์ธ <Bernoulli Process>์˜ ๊ฒฝ์šฐ, Sample Sapce๋Š” 0-1์˜ infinite sequence๊ฐ€ ๋œ๋‹ค.

Property. Sample Space of Bernoulli Process

\[\Omega_{\text{BP}} = \left\{ (b_1, \dots, b_n, \dots ) \mid b_i \in \{ 0, 1 \} \right\}\]

Some Properties of Random Process

<Random Process>๋Š” ์•„๋ž˜์™€ ๊ฐ™์€ ๋ช‡๊ฐ€์ง€ ํŠน์ง•์„ ๊ฐ€์งˆ ์ˆ˜ ์žˆ๋‹ค. ์ด๊ฒƒ์€ ํ•„์ˆ˜์ ์ธ ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ๋ช‡๋ช‡ <random process>์— ๊ณตํ†ต์ ์œผ๋กœ ๋ณด์ด๋Š” ํŠน์ง•์ด๊ฑฐ๋‚˜ ๊ฐ€์ •์ด๋‹ค.

1. Independence btw trials

๊ฐœ๋ณ„ trials์€ ์„œ๋กœ ๋…๋ฆฝ์ ์œผ๋กœ ์ด๋ฃจ์–ด์ง„๋‹ค. ์ฆ‰, ์˜ํ–ฅ์„ ์ฃผ์ง€ ์•Š๋Š”๋‹ค.

  • Bernoulli Process, Poisson Process, โ€ฆ

2. Memoeryless Property

\[P(X = x + k \mid x > k) = P(X = x)\]
  • Bernoulli Process, Poisson Process, โ€ฆ
  • Markov Process

Bernoulli Process (2)

์ด๋ฒˆ ๋ฌธ๋‹จ์—์„œ๋Š” <Bernoulli Process>์— ๋Œ€ํ•œ ๋‚ด์šฉ์—์„œ ์ถ”๊ฐ€์ ์ธ ์ฃผ์ œ๋“ค์„ ๋‹ค๋ฃฌ๋‹ค. ์•„์ง <Bernoulli Process>๊ฐ€ ๋ญ”์ง€ ๋ชจ๋ฅธ๋‹ค๋ฉด, ์œ„์˜ ํฌ์ŠคํŠธ๋ฅผ ๋จผ์ € ์ฝ์–ด๋ณด์ž!

<Bernoulii Process>์—์„œ ์–ด๋–ค random variable $Y$๋ฅผ ์กฐ๊ฑด๊ณผ ํ•จ๊ป˜ ์ •์˜ํ•˜๋ฉด ์ƒˆ๋กœ์šด ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์œ ๋„ํ•  ์ˆ˜ ์žˆ๋‹ค! ์šฐ๋ฆฌ๋Š” <Binomial distribution>, <Geometric distribution>, <Negative BIN distribution>์„ <Bernoulli Process>๋กœ๋ถ€ํ„ฐ ์œ ๋„ํ•ด๋ณด๊ฒ ๋‹ค ๐Ÿ˜

1. Number of Success $S_n$ in $n$ trials.

Letโ€™s derive a random variable $S_n = X_1 + \cdots + X_n$ from the Bernoulli Process.

Then, $S_n$ follows the <Binomial Distribution>!

\[P(S_n = x) = \binom{n}{x} p^x (1-p)^{n-x} \quad \text{for} \; x=0, 1, \dots, n\]

2. Time until the first success

Letโ€™s derive a randome variable $T_1 = \min \{ i \in \mathbb{N} : X_i = 1\}$ from the Bernoulli Process.

Then, $T_1$ follows the <Geometric Distribution>!

\[P(T_1 = x) = P(\underbrace{0, 0, \dots, 0}_{x-1}, 1) = (1-p)^{x-1} p \quad \text{for} \; x=1, 2, \dots\]

3. Time until the first $k$ success

<Geometric Random Variable>์ธ $T_1$์„ ํ™•์žฅํ•œ ๊ฐœ๋…์ด๋‹ค.

Letโ€™s derive a randome variable $T_k = \min \{ i \in \mathbb{N} : | \{ X_i : X_i = 1 \} | = k\}$ from the Bernoulli Process.

Then, $T_n$ follows the <Negative Binomial Distribution>!

\[P(T_k = x) = P(\underbrace{0, 1, \dots, 1, \dots, 0}_{k-1 \text{ success}}, 1) = \binom{x-1}{k-1} (1-p)^{x-k} p^k \quad \text{for} \; x=k, k+1, \dots\]

Poisson Process

<Poisson Process>๋Š” <Bernoulli Process>์—์„œ ๊ทนํ•œ์„ ์ทจํ•ด time interval์˜ ๊ฐ„๊ฒฉ์„ ์•„์ฃผ์•„์ฃผ ์ค„์—ฌ์„œ continuous domain ์œ„์—์„œ ์ •์˜ํ•œ Random Process์ด๋‹ค. BP๊ฐ€ $\mathbb{N}$ ์œ„์—์„œ ์ •์˜๋˜์—ˆ๋‹ค๋ฉด, PP๋Š” $\mathbb{R^{+}}$ ์œ„์—์„œ ์ •์˜๋˜๋Š” Random Process์ธ ์…ˆ!

PP์— ๋Œ€ํ•œ ๋‚ด์šฉ์€ ์•„๋ž˜ ํฌ์ŠคํŠธ์˜ ๋‚ด์šฉ์œผ๋กœ ๋Œ€์ฒดํ•œ๋‹ค ๐Ÿ™

๐Ÿ‘‰ Poisson Process


Gaussian Process

A sequence of Gaussian distribution์œผ๋กœ, multi-variate Gaussian distribution์˜ ์ผ๋ฐ˜ํ™”๋œ ๋ฒ„์ „์ด๋‹ค. โ€œdistribution over functionsโ€์œผ๋กœ ์ทจ๊ธ‰ํ•œ๋‹ค! ๐Ÿ’ช

๐Ÿ‘‰ Distribution over functions & Gaussian Process


Markov Process

๐Ÿ‘‰ Markov Process


references