Sign In

In-Context Black-Box Optimization with Unreliable Feedback

์ž‘์„ฑ์ž
  • Haebom
์นดํ…Œ๊ณ ๋ฆฌ
Empty

์ €์ž

Nicolas Samuel Blumer, Julien Martinelli, Samuel Kaski

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ์—ฐ๊ตฌ๋Š” ๊ณผํ•™ ๋ฐ ๊ณตํ•™ ๋ถ„์•ผ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ธ”๋ž™๋ฐ•์Šค ์ตœ์ ํ™” ๋ฌธ์ œ์—์„œ ์œ ์šฉํ•œ ๋ณด์กฐ ์ •๋ณด(์ „๋ฌธ๊ฐ€, ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ, ์˜ˆ์ธก๊ธฐ, ํœด๋ฆฌ์Šคํ‹ฑ ๋“ฑ)๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ์ด๋Ÿฌํ•œ ๋ณด์กฐ ์ •๋ณด๊ฐ€ ํŽธํ–ฅ๋˜๊ฑฐ๋‚˜ ์ž˜๋ชป๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•˜์—ฌ, ์‚ฌ์ „ ํ•™์Šต๋œ ์˜ตํ‹ฐ๋งˆ์ด์ €๊ฐ€ ์ตœ์ ํ™” ์ด๋ ฅ๊ณผ ํ•จ๊ป˜ ํ˜„์žฌ ํ›„๋ณด ์ง‘ํ•ฉ์— ๋Œ€ํ•œ ์ €๋น„์šฉ ๋ณด์กฐ ํ”ผ๋“œ๋ฐฑ์„ ์กฐ๊ฑด์œผ๋กœ ํ™œ์šฉํ•˜๋Š” "In-context Black-Box Optimization with Unreliable Feedback (FICBO)"๋ฅผ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. FICBO๋Š” ํ”ผ๋“œ๋ฐฑ ์†Œ์Šค์˜ ์ ‘๊ทผ์„ฑ, ๊ด€๋ จ์„ฑ, ์™œ๊ณก ์ •๋„๋ฅผ ๋ชจ๋ธ๋งํ•˜๋Š” ๊ตฌ์กฐํ™”๋œ ํ”ผ๋“œ๋ฐฑ ์‚ฌ์ „(structured feedback prior)์„ ํ†ตํ•ด ํ…Œ์ŠคํŠธ ์‹œ์ ์— ์†Œ์Šค ์‹ ๋ขฐ๋„๋ฅผ ์ถ”์ •ํ•˜๊ณ  ์ฟผ๋ฆฌ ์„ ํƒ์„ ๊ฐœ์„ ํ•˜์—ฌ, ์ •๋ณด์„ฑ ํ”ผ๋“œ๋ฐฑ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ด์šฉํ•˜๊ณ  ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๋Š” ์†Œ์Šค์— ๋Œ€ํ•ด์„œ๋Š” ๊ฐ•๊ฑดํ•จ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
๋‹ค์–‘ํ•œ ์‹ ๋ขฐ๋„๋ฅผ ๊ฐ€์ง„ ์—ฌ๋Ÿฌ ๋ณด์กฐ ํ”ผ๋“œ๋ฐฑ ์†Œ์Šค๋ฅผ ๋™์‹œ์— ๊ณ ๋ คํ•˜์—ฌ ๋ธ”๋ž™๋ฐ•์Šค ์ตœ์ ํ™” ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
ํ…Œ์ŠคํŠธ ์‹œ์ ์˜ ์‹ค์ œ ๋ชฉํ‘œ ํ•จ์ˆ˜ ๊ฐ’๊ณผ ๋ณด์กฐ ์‹ ํ˜ธ๋ฅผ ๋น„๊ตํ•˜์—ฌ ๊ฐ ํ”ผ๋“œ๋ฐฑ ์†Œ์Šค์˜ ์‹ ๋ขฐ๋„๋ฅผ ๋™์ ์œผ๋กœ ์ถ”์ •ํ•˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ๋ชจ๋ธ์˜ ํ•ด์„ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์—ฌ์ค๋‹ˆ๋‹ค.
โ€ข
๊ตฌ์กฐํ™”๋œ ํ”ผ๋“œ๋ฐฑ ์‚ฌ์ „ ๋ฐ ํ”ผ๋“œ๋ฐฑ-์–ด์›จ์–ด ํŠธ๋žœ์Šคํฌ๋จธ ์‚ฌ์ „ ํ•™์Šต์— ๋Œ€ํ•œ ๋” ๊นŠ์€ ํƒ๊ตฌ์™€ ๋‹ค์–‘ํ•œ ์‹ค์ œ ๋ฌธ์ œ์— ๋Œ€ํ•œ ํ™•์žฅ์„ฑ์ด ํ–ฅํ›„ ๊ณผ์ œ๋กœ ๋‚จ์•„์žˆ์Šต๋‹ˆ๋‹ค.
๐Ÿ‘
Made with Slashpage