RL adalah cabang ML yang fokus pada sequential decision making.

Framework

Agent -> action -> Environment -> state + reward
  ^                                      |
  +--------------------------------------+

Key Algorithms

AlgorithmTypeKey Feature
DQNValue-basedExperience replay + target network
PPOPolicy gradientClipped surrogate objective
SACActor-criticMaximum entropy, continuous action
TD3Actor-criticTwin critics, target smoothing
AlphaZeroSelf-playMCTS + neural network

Applications

  • Robotics: manipulation, locomotion, grasping
  • Game AI: AlphaGo, Dota 2
  • Security: adversarial RL for penetration testing