Reinforcement Learning

Definisi

Reinforcement learning (RL) adalah paradigma ML di mana agen belajar dengan berinteraksi dengan environment, menerima reward atau punishment.

Komponen

• Agent — pengambil keputusan
• Environment — dunia tempat agen berada
• State — representasi situasi
• Action — pilihan agen
• Reward — feedback skalar
• Policy — strategi agen

Algoritma

• Q-learning (1989)
• DQN (Deep Q-Network, 2013) — DeepMind
• Policy Gradient (REINFORCE, 1992)
• Actor-Critic (A2C, A3C)
• PPO (Proximal Policy Optimization, 2017) — OpenAI
• GRPO (Group Relative Policy Optimization, 2024) — DeepSeek

Aplikasi

• Game AI — AlphaGo, AlphaZero, OpenAI Five (Dota 2)
• Robotics — locomotion, manipulation
• RLHF — alignment LLM
• Autonomous driving