ExploRLLM: Guiding Exploration in Reinforcement Learning with Large Language Models

์ €์ž: Runyu Ma, Jelle Luijkx, Zlatan Ajanovic, Jens Kober | ๋‚ ์งœ: 2024-03-14 | URL: https://arxiv.org/abs/2403.09583 📄 PDF


Essence

Figure 1

Fig. 1: Graphical overview of ExploRLLM.

ExploRLLM์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด ์ƒ์„ฑํ•œ ์ •์ฑ… ์ฝ”๋“œ๋กœ RL ์—์ด์ „ํŠธ์˜ ํƒ์ƒ‰์„ ์œ ๋„ํ•˜๋ฉด์„œ, ์ž”์ฐจ RL ์—์ด์ „ํŠธ๊ฐ€ FM์˜ ๋ฌผ๋ฆฌ์  ์ดํ•ด ๋ถ€์กฑ์„ ๋ณด์™„ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ๋กœ๋ด‡ ์กฐ์ž‘ ์ž‘์—…์˜ ์ƒ˜ํ”Œ ํšจ์œจ์„ฑ๊ณผ ์ˆ˜๋ ด์„ฑ์„ ๊ฐœ์„ ํ•œ๋‹ค.

Motivation

Achievement

Figure 4

Fig. 4: Training curves for varying exploration rates in SH and LH tasks. ExploRLLM outperforms the exploration policies

How

Figure 2

Fig. 2: Implementation structure of ExploRLLM for tabletop manipulation, combining the strengths of RL and FMs.

Originality

Limitation & Further Study

Evaluation

Novelty: 4/5 Technical Soundness: 3/5 Significance: 4/5 Clarity: 4/5 Overall: 4/5

์ดํ‰: ExploRLLM์€ FM๊ณผ RL์˜ ์žฅ์ ์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ฒฐํ•ฉํ•˜์—ฌ ๋กœ๋ด‡ ์กฐ์ž‘์˜ ์ƒ˜ํ”Œ ํšจ์œจ์„ฑ์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•˜๋Š” ์‹ค์šฉ์ ์ธ ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•˜๋ฉฐ, ํŠนํžˆ LLM ๊ธฐ๋ฐ˜ ํƒ์ƒ‰ ์ „๋žต์˜ ํ˜์‹ ์„ฑ๊ณผ ์‹ค์ œ ๋กœ๋ด‡์—์„œ์˜ zero-shot ์ „์ด ์„ฑ๊ณต์€ ๋†’์€ ๊ฐ€์น˜๋ฅผ ๊ฐ€์ง„๋‹ค. ๋‹ค๋งŒ ํ‰๊ฐ€ ๋ฒ”์œ„ ํ™•๋Œ€์™€ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ ๊ฒ€์ฆ์ด ํ•„์š”ํ•˜๋‹ค.

← ๋ชฉ๋ก์œผ๋กœ ๋Œ์•„๊ฐ€๊ธฐ

๐ŸŽง Audio Overview

์ด ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ๋ฅผ ํŒŸ์บ์ŠคํŠธํ˜• ์˜ค๋””์˜ค๋กœ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. (Gemini ยท ํ‚ค๋Š” ๋ธŒ๋ผ์šฐ์ €์—๋งŒ ์ €์žฅ ยท ์™„์„ฑ๋ณธ์€ ์ด๋ฉ”์ผ๋กœ๋„ ์ „์†ก)
โ–ธ ๊ณ ๊ธ‰: ๊ตฌ์„ฑ ๋ฐฉํ–ฅ(๋Œ€๋ณธ ์ž‘์„ฑ ์ง€์นจ) ์ง์ ‘ ์ˆ˜์ •