λ°μν
Agentic AI μμ€ν
μμ κ° Unit-Agentμ μ€ν μ λ΅(μ: RAG κ²μ κΉμ΄, λͺ¨λΈ μ ν, ν둬ννΈ ν¨ν΄)μ
Reinforcement Learning (RL) κΈ°λ°μΌλ‘ μ€μ€λ‘ μ΅μ ννλλ‘ μ€κ³νλ € νλ€.
μ 쑰곡μ λ³΄κ³ μ μλν μλ리μ€λ₯Ό κΈ°μ€μΌλ‘,
μν(State), νλ(Action), 보μ(Reward)μ μ μνκ³ ,
κ°ννμ΅μ΄ Agent μμ¬κ²°μ κ°μ μ μ΄λ»κ² κΈ°μ¬νλμ§ μμ νμμ€.
β κ°μ
- κΈ°μ‘΄ Agentic AIλ μ ν΄μ§ ν둬ννΈ μ²΄μΈμ λ°λΌ μ€ννμ§λ§,
μ€μ λ³΄κ³ μ νμ§μ νκ²½(λ°μ΄ν°λ, 곡μ 쑰건, LLMμλ΅ νμ§) μ λ°λΌ λ¬λΌμ§λ€. - λ°λΌμ κ° μ€ν 루νμμμ μ ν(action) μ κ°ννμ΅μΌλ‘ μ‘°μ νμ¬
μ€μ€λ‘ μ΅μ μ μ λ΅μ νμ΅νλλ‘ λ§λ λ€.
β‘ κ°ννμ΅ κ΅¬μ±μμ μ μ
ꡬμ±μμλ΄μ©μ 쑰곡μ μμ
| State (S) | νμ¬ μμ€ν μ λ΄λΆ μν (μ λ ₯ κΈΈμ΄, μ λ’°λ, RAG κ²°κ³Ό μ, μ΄μ μ μ) | {context_len=1800, conf=0.85, retries=1} |
| Action (A) | Agentκ° μ ν κ°λ₯ν μ€ν μ΅μ | {RAG depth=2, Model=TGI-v3, Prompt=‘formal’} |
| Reward (R) | νκ° μ§ν κΈ°λ° λ³΄μ ν¨μ | R = 0.4·Groundedness + 0.4·Citation + 0.2·TaskSuccess |
β’ RL 루ν λμ ꡬ쑰
νκ²½(Environment)
↑ β
β ↓
Agent(Planner)
→ μ ν(Action): λͺ¨λΈ·κ²μμ 롷체μΈκ²½λ‘
→ μ€ν ν νκ°(Self-Eval Harness)
→ 보μ κ³μ°(R)
→ Policy μ
λ°μ΄νΈ(θ ← θ + α·∇θ)
- Algorithm: PPO (Proximal Policy Optimization) λλ DQN κΈ°λ°
- Training Signal: Self-Eval Harnessκ° μ 곡νλ G, C, T μ μ → Reward
β£ μμ μλ리μ€
Episodeμ νλ μ λ΅G/C/T보μ(R)
| #1 | RAG depth=1, model=vLLM7B | 0.72/0.69/0.9 | 0.70 |
| #2 | RAG depth=3, model=Exaone13B | 0.93/0.95/1.0 | 0.94 |
| #3 | RAG depth=4, model=Exaone13B | 0.94/0.91/1.0 | 0.92 |
→ μ΅μ μ μ± : “RAG depth=3, model=Exaone13B, style=formal” μ ν νλ₯ ↑
β€ Reward Engineering
- λ¨μΌ μ§νκ° μλ μ‘°ν©ν 보μ μ¬μ©:
- 보λμ€/ν¨λν°:
- +0.1 → if response < 3s
- −0.2 → if hallucination_detected=True
R = 0.4 * groundedness + 0.4 * citation + 0.2 * task_success
β₯ μ€λ¬΄μ ν¨κ³Ό
νλͺ©κΈ°μ‘΄RL κΈ°λ°
| μ€νμ λ΅ | μλ μ€μ | μλ μ μ |
| νμ§ λ³λ | νΌ | μμ ν |
| λΉμ© ν¨μ¨ | μΌμ | μ±λ₯ λλΉ μ΅μ ν |
| Self-Learning | μμ | μ‘΄μ¬ |
β¦ νκ° ν¬μΈνΈ
- μν·νλ·λ³΄μ μ μλ₯Ό μ νν κΈ°μ νλκ°
- 보μμμ Groundedness·Citationμ μ°κ²°νλκ°
- RLμ΄ Agent μ±λ₯μ κ°μ νλ λ Όλ¦¬λ₯Ό μ μνλκ°
λ°μν