greentec's blog game designer, scripter, researcher

en kr

CoG 2019 ์ฐธ์„ ํ›„๊ธฐ

Tags:


์ง€๋‚œ ์ฃผ์— ๋Ÿฐ๋˜ QMUL ๋Œ€ํ•™์—์„œ ๊ฐœ์ตœ๋˜์—ˆ๋˜ CoG(Conference on Games) 2019 ์— ๋…ผ๋ฌธ์ด Accept๋˜์–ด ์ฐธ์„ํ•˜๊ณ  ๋Œ์•„์™”์Šต๋‹ˆ๋‹ค. ์ƒ์•  ์ฒซ ํ•™ํšŒ ์ฐธ์„์— ์˜์–ด ๋ฐœํ‘œ๋„ ์ฒ˜์Œ์ด๋ผ ๊ธด์žฅ์ด ๋งŽ์ด ๋˜์—ˆ์ง€๋งŒ ์ผ๋‹จ ์ค€๋น„ํ•œ ๋‚ด์šฉ์€ ๋‹ค ํ•˜๊ณ  ๋Œ์•„์™”์Šต๋‹ˆ๋‹ค. ๊ฐœ์ธ์ ์œผ๋กœ๋Š” ์„ธ๊ณ„์ ์ธ ์—ฐ๊ตฌ์ž๋“ค์˜ ๋ฐœํ‘œ๋ฅผ ๋ณด๊ณ  ์ž๊ทน์ด ๋งŽ์ด ๋˜๋Š” ์‹œ๊ฐ„์ด์—ˆ๊ธฐ ๋•Œ๋ฌธ์—, ๊ธฐ์–ต์ด ํฌ๋ฏธํ•ด์ง€๊ธฐ ์ „์— ์ฐธ์„ ํ›„๊ธฐ๋ฅผ ๋‚จ๊ฒจ๋ด…๋‹ˆ๋‹ค.

ย 

Conference on Games

2018๋…„๊นŒ์ง€ CIG(Computational Intelligence and Games)์˜€๋˜ ์ด ํ•™ํšŒ๋Š” ์ข€ ๋” ๋ฒ”์œ„๋ฅผ ๋„“ํžˆ๊ณ ์ž 2019๋…„๋ถ€ํ„ฐ Conference on Games๋ผ๋Š” ์ด๋ฆ„์œผ๋กœ ๋ฐ”๋€Œ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฃจ๊ณ  ์žˆ๋Š” ์ฃผ์ œ๋กœ๋Š” ์•„์ง๊นŒ์ง€๋Š” AI์™€ PCG ์ชฝ์˜ ๋…ผ๋ฌธ์ด ๋‹ค์ˆ˜๋ฅผ ์ฐจ์ง€ํ•˜์ง€๋งŒ ๊ทธ ์™ธ์—๋„ ํ”Œ๋ ˆ์ด์–ด ๋ชจ๋ธ๋ง, serious games, ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ, e-sports ๋“ฑ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์˜ ๋…ผ๋ฌธ๊ณผ ๋ฐœํ‘œ๋ฅผ ์ฐพ์•„๋ณผ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.

๊ฒŒ์ž„ ์ชฝ์˜ ํ•™ํšŒ๋กœ๋Š” ๋ฏธ๊ตญ ์„œ๋ถ€๊ถŒ ์ค‘์‹ฌ์˜ AIIDE(AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment)๊ฐ€ ์žˆ๋‹ค๋ฉด ๋ฏธ๊ตญ ๋™๋ถ€์™€ ์œ ๋Ÿฝ์˜ ์—ฐ๊ตฌ์ž๋“ค์ด ์ค‘์‹ฌ์ด ๋˜๋Š” CoG๋ฅผ ์–‘๋Œ€ ํ•™ํšŒ๋กœ ๊ผฝ์„ ์ˆ˜ ์žˆ๊ฒ ์Šต๋‹ˆ๋‹ค.

์ฃผ์š” ์—ฐ๊ตฌ์ž๋กœ๋Š” ๋‰ด์š• ๋Œ€ํ•™์˜ Julian Togelius ๊ต์ˆ˜, ๋ชฐํƒ€ ๋Œ€ํ•™์˜ Georgios N. Yannakakis ๊ต์ˆ˜, ๋„ค๋œ๋ž€๋“œ ๋งˆ์ŠคํŠธ๋ฆฌํํŠธ ๋Œ€ํ•™์˜ Cameron Browne ๊ต์ˆ˜(Evolutionary Search๋กœ ๋งŒ๋“  ๋ณด๋“œ ๊ฒŒ์ž„์ธ Yavalath๋กœ ์ž˜ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค), ์ฝ”ํŽœํ•˜๊ฒ ๋Œ€ํ•™์˜ Sebastian Risi ๊ต์ˆ˜ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.

๋‘˜์งธ ๋‚  ์žˆ์—ˆ๋˜ ์ Š์€ ์—ฐ๊ตฌ์ž๋“ค์˜ ๋ชจ์ž„. ์ €๋Š” ์–‘์‹ฌ์ƒ ๊ฐ€์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์‚ฌ์ง„ ์ถœ์ฒ˜ ํŠธ์œ— ๋งํฌ

ย 

๋‹ค์–‘์„ฑ

CoG2019์—์„œ ๋Š๋‚€ ์  ์ค‘ ํ•˜๋‚˜๋Š” ๋‹ค์–‘์„ฑ์ด์—ˆ์Šต๋‹ˆ๋‹ค. Angry birds, ์Šคํƒ€ํฌ๋ž˜ํ”„ํŠธ, ๋งˆ์ธํฌ๋ž˜ํ”„ํŠธ ๋“ฑ ์ด 12๊ฐœ ์ข…๋ชฉ์˜ competition์ด ์—ด๋ ธ๊ณ  ๊ฐ๊ฐ์— ๋Œ€ํ•œ ์ˆ˜์ƒ์ž ๋ฐœํ‘œ๊ฐ€ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ €๋„ ๋งˆ์นจ 1ํšŒ์ฐจ์ธ bot-bowl competition์— ์ฐธ๊ฐ€ํ•ด์„œ ์šด์ด ์ข‹๊ฒŒ 2๋“ฑ์œผ๋กœ ์ž…์ƒํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ competition์ด ๋„ˆ๋ฌด ๋งŽ๋‹ค๋ณด๋‹ˆ ๊ฐ competition์— ๋Œ€ํ•œ ์ฐธ์—ฌ์ž ์ˆ˜๊ฐ€ ์ ์€ ๊ฒƒ์ด ๋ฌธ์ œ ์ œ๊ธฐ๊ฐ€ ๋˜์–ด ์ˆ˜์ƒ์ž ๋ฐœํ‘œ ์‹œ๊ฐ„ ๋„์ค‘์— ๊ทธ์— ๋Œ€ํ•œ ํŒจ๋„ ํ† ๋ก ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. competition์˜ ์ˆ˜๋ฅผ ์ค„์ด๊ณ  ์ง‘์ค‘ํ•˜์ž๋Š” ์˜๊ฒฌ๋„ ์žˆ์—ˆ๊ณ , ๋‹ค์–‘์„ฑ์„ ์œ ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ competition์„ ๊ฐœ์ตœํ•˜๋Š” ๊ฒƒ์„ ๋ง‰์•„์„œ๋Š” ์•ˆ๋œ๋‹ค๋Š” ์˜๊ฒฌ๋„ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ €๋Š” ๊ฐœ์ธ์ ์œผ๋กœ ํ›„์ž์— ๋™์˜ํ•˜๋Š” ํŽธ์ด์—ˆ์Šต๋‹ˆ๋‹ค๋งŒ, ์ด๋ ‡๊ฒŒ ๋Œ€ํšŒ์˜ ์ง„ํ–‰ ๋ฐฉ์‹์— ๋Œ€ํ•œ ์ž์œ ๋กœ์šด ๋…ผ์˜๊ฐ€ ์—ด๋ฆฌ๋Š” ๊ฒƒ ์ž์ฒด๊ฐ€ ์ธ์ƒ์ ์ด์—ˆ์Šต๋‹ˆ๋‹ค.

๋” ๋‚˜์€ competition์„ ์œ„ํ•œ ํŒจ๋„ ํ† ๋ก . ์‚ฌ์ง„ ์ถœ์ฒ˜ ํŠธ์œ— ๋งํฌ

๋‹ค์–‘์„ฑ์€ ๋…ผ๋ฌธ ์—ฐ๊ตฌ ๋ถ„์•ผ์—์„œ๋„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๊ฐ€์žฅ ์ฃผ๋ชฉ์„ ๋ฐ›๊ณ  ์žˆ๋Š” ๋”ฅ๋Ÿฌ๋‹๊ณผ ๊ฐ•ํ™”ํ•™์Šต ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ง„ํ™”์  ํƒ์ƒ‰(Evolutionary Search), ์…€๋ฃฐ๋Ÿฌ ์˜คํ† ๋งˆํƒ€(Cellular Automata), ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•œ problem solving ๋“ฑ ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์— ๋Œ€ํ•œ ๋…ผ๋ฌธ์ด ๋ฐœํ‘œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

๋ฐœํ‘œ๋ฅผ ์ง์ ‘ ๋“ค์€ ๋…ผ๋ฌธ ์ค‘ ๋‘ ๊ฐœ๋ฅผ ๋ฝ‘์•„์„œ ๊ฐ„๋‹จํžˆ ์†Œ๊ฐœํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ๋จผ์ € ๋งˆ์ŠคํŠธ๋ฆฌํํŠธ ๋Œ€ํ•™์˜ Jos Uiterwijk ๊ต์ˆ˜๊ฐ€ ๋ฐœํ‘œํ•œ โ€œSolving Strong and Weak 4-in-a-Rowโ€์ž…๋‹ˆ๋‹ค. 4๋ชฉ ๊ฐ™์€ ํ„ด์ œ ๋ณด๋“œ ๊ฒŒ์ž„์€ ์„œ๋กœ ์ตœ์ (optimal)์˜ ์ˆ˜๋ฅผ ๋‘˜ ๊ฒฝ์šฐ ์ฒซ๋ฒˆ์งธ ํ”Œ๋ ˆ์ด์–ด๊ฐ€ ๋ฌด์กฐ๊ฑด ์œ ๋ฆฌํ•œ๋ฐ, 5x6 ๋ณด๋“œ์™€ 4x9 ๋ณด๋“œ ์ด์ƒ์—์„œ ์ฒซ๋ฒˆ์งธ ํ”Œ๋ ˆ์ด์–ด(ํ‘)์ด ๋จผ์ € ๋‘˜ ๊ฒฝ์šฐ ๋ฌด์กฐ๊ฑด ์ด๊ธฐ๊ฒŒ ๋œ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ ์œผ๋กœ ์ฆ๋ช…ํ•œ ๋…ผ๋ฌธ์ž…๋‹ˆ๋‹ค.

5x6 ๋ณด๋“œ์™€ 4x9 ๋ณด๋“œ ์ด์ƒ์—์„œ ์ฒซ๋ฒˆ์งธ ํ”Œ๋ ˆ์ด์–ด(ํ‘)์ด ๋จผ์ € ๋‘˜ ๊ฒฝ์šฐ ์„œ๋กœ ์ตœ์ ์˜ ์ˆ˜๋ฅผ ๋‘˜ ๋•Œ ๋ฌด์กฐ๊ฑด ์ด๊ธฐ๋Š” ์ˆ˜๊ฐ€ ๋ฐœ๊ฒฌ๋ฉ๋‹ˆ๋‹ค.

์˜คํ•˜์ด์˜ค ์ฃผ๋ฆฝ๋Œ€ํ•™์˜ Pavlos Androulakakis๊ฐ€ ๋ฐœํ‘œํ•œ โ€œEvolution of Kiting Behavior in a Two Player Combat Problemโ€๋„ ์ƒˆ๋กœ์› ์Šต๋‹ˆ๋‹ค. ์ง„ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‘ ํ”Œ๋ ˆ์ด์–ด์˜ ์ „ํˆฌ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์—์„œ์˜ ์ตœ์  control์„ ์ฐพ๋Š” ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•œ ์ด ๋…ผ๋ฌธ์€ ์ตœ์ ์ด๋ผ๊ณ  ์ƒ๊ฐํ•˜๋Š” ์›€์ง์ž„๋ณด๋‹ค grid parameterization์„ ์ˆ˜ํ–‰ํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ง„ํ™”์ ์œผ๋กœ ํƒ์ƒ‰ํ–ˆ์„ ๋•Œ ํผํฌ๋จผ์Šค๊ฐ€ ๋” ๋‚˜์•„์งˆ ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค.

grid parameterization์œผ๋กœ ์ง„ํ™”์  ํƒ์ƒ‰์„ ์ˆ˜ํ–‰ํ•˜๋ฉด 2D State Space Control์— ๊ฐ€๊นŒ์›Œ์ง€๋Š” ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์˜ค๊ณ , ํผํฌ๋จผ์Šค๊ฐ€ ๋” ๋‚˜์•„์งˆ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.

ย 

ํ”„๋ ˆ์ž„์›Œํฌ

CoG2019์—์„œ๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค๋ ค๋Š” ๋ช‡๋ช‡ ์‹œ๋„๊ฐ€ ๋ˆˆ์— ๋„์—ˆ์Šต๋‹ˆ๋‹ค. ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ํŠน์ • ์—ฐ๊ตฌ์—์„œ ๋ฒ—์–ด๋‚˜ ๋‹ค์–‘ํ•˜๊ณ  ์ผ๋ฐ˜์ ์ธ ์—ฐ๊ตฌ๋ฅผ ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ฃผ๊ณ , ๋‹ค๋ฅธ ์—ฐ๊ตฌ์ž๋“ค์„ ์ด ํ”„๋ ˆ์ž„์›Œํฌ์— ๋Œ์–ด๋“ค์—ฌ์„œ ์ „์ฒด์ ์ธ ์—ฐ๊ตฌ์˜ ์งˆ์„ ๋†’์ด๋Š” ํšจ๊ณผ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.

์ง„ํ™”์  ํƒ์ƒ‰์œผ๋กœ Yavalath๋ฅผ ๋งŒ๋“  Cameron Browne ๊ต์ˆ˜์˜ ํŒ€์—์„œ๋Š” 4๊ฐœ์˜ ๋…ผ๋ฌธ์„ ๋ฐœํ‘œํ–ˆ๋Š”๋ฐ, ์ด ์ค‘ ์•ผ์‹ฌ์ฐจ๊ฒŒ ๋ฐœํ‘œํ•œ Ludii Portal์— ๋Œ€ํ•œ ๋…ผ๋ฌธ(โ€œLudii as a Competition Platformโ€)์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. Yavalath๋ฅผ ๋งŒ๋“  ๊ฒƒ์ฒ˜๋Ÿผ ๊ฐ„๋‹จํ•œ ๊ทœ์น™์œผ๋กœ ๊ฒŒ์ž„์„ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์„ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ, ๊ฒŒ์ž„์˜ ๋””์ž์ธ์  ์ธก๋ฉด์„ ๋ถ„์„, ํ‰๊ฐ€ํ•˜๊ณ  ๊ฐ ๊ฒŒ์ž„์— ๋Œ€ํ•œ ์˜จ๋ผ์ธ ๋งค์น˜๋ฉ”์ดํ‚น์„ ์ œ๊ณตํ•˜๋Š” ํ”Œ๋žซํผ์ž…๋‹ˆ๋‹ค. ์•„์ง ๊ฐœ๋ฐœ์ด ์™„๋ฃŒ๋˜์ง€ ์•Š์€ ๋ถ€๋ถ„๋“ค์ด ์žˆ์ง€๋งŒ ๊ฒŒ์ž„ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ์ข‹์€ ์‹œ๋„๋ผ๊ณ  ์ƒ๊ฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ์— ๋”ฅ๋งˆ์ธ๋“œ์—์„œ ๋ฐœํ‘œํ•œ OpenSpiel๋„ 21๊ฐœ์˜ ๊ฐ„๋‹จํ•œ ๋ณด๋“œ ๊ฒŒ์ž„์— ๋Œ€ํ•œ ์—ฐ๊ตฌ/๊ฐ•ํ™”ํ•™์Šต ํ™˜๊ฒฝ์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ ๋น„์Šทํ•œ ์‹œ๋„๋ฅผ ํ•˜๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค.

์™ผ์ชฝ์€ Yavalath์˜ ๊ทœ์น™, ์˜ค๋ฅธ์ชฝ์€ ๊ฒŒ์ž„ ํ™”๋ฉด์ž…๋‹ˆ๋‹ค. 4๋ชฉ์ด์ง€๋งŒ ๊ทธ ์ „์— 3๋ชฉ์„ ๋งŒ๋“ค๋ฉด ํŒจ๋ฐฐํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.

์ฝ”ํŽœํ•˜๊ฒ ๋Œ€ํ•™์˜ Mads Johansen์€ GVGAI Competition์—์„œ ์‚ฌ์šฉ๋œ ๊ฒƒ์œผ๋กœ ์œ ๋ช…ํ•œ VGDL(Video Game Description Language)1์„ ์œ ๋‹ˆํ‹ฐ์˜ ML-Agents์— ํ†ตํ•ฉํ•œ ํ•™์Šต ํ™˜๊ฒฝ์„ ๊ตฌ์ถ•ํ•œ ๋…ผ๋ฌธ(โ€œVideo Game Description Language Environment for Unity Machine Learning Agentsโ€)์„ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค. ์œ ๋‹ˆํ‹ฐ์˜ ML-Agents๋Š” PPO ๊ฐ™์€ ์ตœ์‹ ์˜ ๊ฐ•ํ™”ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํด๋ฆญ ํ•œ๋ฒˆ์œผ๋กœ ์‰ฝ๊ฒŒ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ์—ฌ๊ธฐ์— ๊ฒŒ์ž„ ํ™˜๊ฒฝ๊นŒ์ง€ ์‰ฝ๊ฒŒ ์„ธํŒ…ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋Š” ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์— ์ผ๋ฐ˜ํ™”๋˜๋ฉด ๋‹ค์–‘ํ•œ ์—ฐ๊ตฌ์— ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ์ƒ๊ฐ๋ฉ๋‹ˆ๋‹ค.

VGDL ๋…ผ๋ฌธ์—์„œ ์ œ์‹œ๋œ zelda-like ๊ฒŒ์ž„์˜ ํ‘œํ˜„๊ณผ ๊ทœ์น™. ์ถœ์ฒ˜ ๋งํฌ

ย 

MCTS์˜ ์•ฝ์ง„

์ด๋ฒˆ CoG2019์˜ Best Paper ์ˆ˜์ƒ ๋ถ€๋ฌธ์—๋Š” ์ด 6๊ฐœ์˜ ํ›„๋ณด๊ฐ€ ์˜ฌ๋ž๋Š”๋ฐ, ๊ทธ ์ค‘ 3๊ฐœ๊ฐ€ MCTS๋ฅผ ํ™œ์šฉํ•œ ๋…ผ๋ฌธ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด์—๋„ ๊ฐ•ํ™”ํ•™์Šต์˜ ๊ฐ•๋ ฅํ•œ ๋ฐฉ๋ฒ•๋ก ์ด์—ˆ๋˜ MCTS๋Š” ๋”ฅ๋งˆ์ธ๋“œ์˜ ์•ŒํŒŒ๊ณ ์™€ ์•ŒํŒŒ์ œ๋กœ์—์„œ ์‚ฌ์šฉ๋œ ์ดํ›„ ์ตœ๊ณ ์˜ ์ฃผ๋ชฉ์„ ๋ฐ›๊ณ  ์žˆ๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๊ฐ ํŽ˜์ดํผ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

โ€œLearning Policies from Self-Play with Policy Gradients and MCTS Value Estimatesโ€, Dennis Soemers, Eric Piette, Matthew Stephenson and Cameron Browne

โ€œRandom Search-Control Parameters in Monte-Carlo Tree Searchโ€, Chiara F. Sironi and Mark H. M. Winands

โ€œMysterious Murder - MCTS-driven Murder Mystery Generationโ€, Corinna Jaschek, Tom Beckmann, Jaime Garcia and William Raffe

์•ž์˜ ๋‘ ๋…ผ๋ฌธ์€ AI์— ๋Œ€ํ•œ ๊ฒƒ์ธ๋ฐ ๋งˆ์ง€๋ง‰์€ ํŠน์ดํ•˜๊ฒŒ๋„ PCG์— MCTS๋ฅผ ํ™œ์šฉํ•œ ์‚ฌ๋ก€์˜€์Šต๋‹ˆ๋‹ค. ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ ์ƒ์„ฑ์— MCTS๋ฅผ ์‚ฌ์šฉํ–ˆ๋‹ค๋Š” ์ ์ด ๊ธ์ •์ ์œผ๋กœ ์ž‘์šฉ๋˜์—ˆ๋Š”์ง€ ์ด ๋…ผ๋ฌธ์€ 3๊ฐœ์˜ Best Paper ์ค‘ ํ•˜๋‚˜๋ฅผ ์ˆ˜์ƒํ–ˆ์Šต๋‹ˆ๋‹ค. CoG2019์— MCTS ๋˜๋Š” Monte-Carlo Tree Search๊ฐ€ ์ œ๋ชฉ์— ๋“ค์–ด๊ฐ„ ๋…ผ๋ฌธ์€ ์ด 9๊ฐœ๋กœ, ์ „์ฒด 161๊ฐœ์˜ ๋…ผ๋ฌธ ์ค‘ ์•ฝ 5.6%์˜€์Šต๋‹ˆ๋‹ค.

ย 

PCG

๊ฒŒ์ž„ ํ•™ํšŒ์—์„œ PCG๋Š” AI์™€ ํ•จ๊ป˜ ํ•ญ์ƒ ๋ฉ”์ธ ํ…Œ๋งˆ ์ค‘ ํ•˜๋‚˜๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์ธ์ƒ์ ์ด์—ˆ๋˜ PCG ๋…ผ๋ฌธ์„ ์—ญ์‹œ 2๊ฐœ ์ •๋„ ๋ฝ‘์•„๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.

์บ๋‚˜๋‹ค Simon Fraser University์˜ Ahmed Abuzuraiq์ด ๋ฐœํ‘œํ•œ โ€œTaksim: A Constrained Graph Partitioning Framework for Procedural Content Generationโ€์€ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ๋ฅผ ๊ณต๊ฐ„์  ๊ตฌ์กฐ(๋งต)๋กœ Answer Set Programming์„ ์‚ฌ์šฉํ•ด์„œ ์˜ฎ๊ธฐ๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ๋‚ด์šฉ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ œ๊ฐ€ ์˜ˆ์ „์— ์—ฐ๊ตฌํ–ˆ๋˜ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ๋ฅผ WFC๋ฅผ ์‚ฌ์šฉํ•ด์„œ ๋งต์— ์˜ฎ๊ธฐ๋Š” ๋ฐฉ๋ฒ•๊ณผ ๋น„์Šทํ•˜๊ธฐ๋„ ํ•ด์„œ ์—ด์‹ฌํžˆ ๋ดค๋˜ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค.

์™ผ์ชฝ์˜ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ๋ฅผ ์˜ค๋ฅธ์ชฝ์˜ ๊ณต๊ฐ„์  ๊ตฌ์กฐ๋กœ ์˜ฎ๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

๋‰ด์š•๋Œ€ํ•™๊ต์˜ Sarjak Pankaj Thakkar๊ฐ€ ๋ฐœํ‘œํ•œ โ€œAutoencoder and Evolutionary Algorithm for Level Generation in Lode Runnerโ€๋Š” ๋ ˆ๋ฒจ ์ƒ์„ฑ์— Variational AutoEncoder์™€ ์ง„ํ™”์  ํƒ์ƒ‰์„ ์ด์šฉํ•ด์„œ ์ข‹์€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ €๋„ 2017๋…„์— Variational AutoEncoder๋กœ ๋ ˆ๋ฒจ์„ ์ƒ์„ฑํ•ด๋ณธ ์ ์ด ์žˆ๋Š”๋ฐ, ์ด๋ ‡๊ฒŒ ์ข‹์€ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์˜ค์ง€๋Š” ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๊ฐœ์ธ์ ์œผ๋กœ ์ž๊ทน์ด ๋˜๊ณ  ์‹œ๋„ํ•ด๋ณด๊ณ  ์‹ถ์€ ๋ฐฉ๋ฒ•๋ก ์ด์—ˆ์Šต๋‹ˆ๋‹ค.

์™ผ์ชฝ์—์„œ AutoEncoder์™€ Variational AutoEncoder๋ฅผ ๋‹จ์ˆœ ์ ์šฉํ•œ ๊ฒฐ๊ณผ์— ๋น„ํ•ด, ์ง„ํ™”์  ํƒ์ƒ‰ ๊ณผ์ •์„ ๊ฑฐ์น˜๋ฉด ์˜ค๋ฅธ์ชฝ์˜ ๊ฒฐ๊ณผ๋Š” ํ›จ์”ฌ ๊น”๋”ํ•ด์ง‘๋‹ˆ๋‹ค.

ย 

๋งˆ์น˜๋ฉฐ

ํ•™ํšŒ์˜ ๋ชจ๋“  ์ผ์ •๊ณผ ๋ฐœํ‘œ๋ฅผ ๋๋‚ด๊ณ  ์ง‘์— ๋Œ์•„์˜ค๋Š” ๋น„ํ–‰๊ธฐ ์•ˆ์—์„œ ๋งŽ์€ ์ƒ๊ฐ์ด ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ๋‚˜๋Š” ์ž˜ํ•˜๊ณ  ์žˆ๋Š”์ง€? ์•ž์œผ๋กœ ์–ด๋–ค ์—ฐ๊ตฌ๋ฅผ ํ•˜๋Š” ์ชฝ์œผ๋กœ ๋ฐฉํ–ฅ์„ ์žก์•„์•ผ ํ•˜๋Š”์ง€? ์„ธ๊ณ„ ๋ฌด๋Œ€์—์„œ ์‚ฌ๋žŒ๋“ค์ด ์–ด๋–ค ์ƒ๊ฐ์„ ํ•˜๊ณ  ์–ด๋–ค ์—ฐ๊ตฌ๋ฅผ ํ•˜๋Š”์ง€์— ๋Œ€ํ•ด์„œ ๋งŽ์ด ๋ฐฐ์šธ ์ˆ˜ ์žˆ๋Š” ์‹œ๊ฐ„์ด์—ˆ๊ณ  ํ•œ๊ตญ 36๋„, ๋Ÿฐ๋˜ 20๋„์˜ ํฐ ๊ธฐ์˜จ์ฐจ๋กœ ๊ฐ€์ž๋งˆ์ž ๊ฐ๊ธฐ์— ๊ฑธ๋ฆฌ๋Š” ๋“ฑ ๋ชธ์€ ํ”ผ๊ณคํ–ˆ์ง€๋งŒ ๋ณด๋žŒ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.

๋‚ด๋…„์˜ CoG2020์€ ์ผ๋ณธ์˜ ์˜ค์‚ฌ์นด์—์„œ ์—ด๋ฆฌ๊ณ  ํ•œ๊ตญ GIST์˜ ๊น€๊ฒฝ์ค‘ ๊ต์ˆ˜๋‹˜๊ณผ ์ผ๋ณธ ๋ฆฌ์ธ ๋ฉ”์ด์นธ ๋Œ€ํ•™์˜ Ruck Thawonmas ๊ต์ˆ˜๋‹˜์ด ๊ณต๋™์œผ๋กœ General Chairs๋ฅผ ๋งก์œผ์…จ์Šต๋‹ˆ๋‹ค. 2011๋…„ ํ•œ๊ตญ ์„œ์šธ, 2015๋…„ ๋Œ€๋งŒ ์ดํ›„๋กœ ์˜ค๋žœ๋งŒ์— ์•„์‹œ์•„์—์„œ ์—ด๋ฆฌ๋Š” CoG๋ผ์„œ ์•„์‹œ์•„์—์„œ๋„ ๋งŽ์€ ๋ถ„๋“ค์ด ์ฐธ์„ํ•˜์‹ค ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์ด ๊ธ€์„ ๋ณด๊ณ  ๊ด€์‹ฌ์ด ์ƒ๊ธฐ์‹  ๋ถ„๋“ค์€ ๋‚ด๋…„์— CoG2020์—์„œ ๋งŒ๋‚˜๋ต ์ˆ˜ ์žˆ์œผ๋ฉด ์ข‹๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

๋‚ด๋…„์— ๋˜ ๊ฐˆ ์ˆ˜ ์žˆ์œผ๋ฉด ์ข‹๊ฒ ์Šต๋‹ˆ๋‹ค! ๊ทธ๋Ÿฌ๋ ค๋ฉด ์ข‹์€ ๋…ผ๋ฌธ์„ ์จ์•ผ๊ฒ ๋„ค์š”. ์‚ฌ์ง„ ์ถœ์ฒ˜ ํŠธ์œ— ๋งํฌ

  1. VGDL์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ์ด ๋…ผ๋ฌธ์— ๋‚˜์™€์žˆ์Šต๋‹ˆ๋‹ค.ย