Method
Reinforcement learning
Value or policy learning (DQN, DDPG) incl. offline RL.
21 cases from 17 companies apply Reinforcement learning. 7 were checked in an online A/B test; 2 report results in production. Includes variants: PPO.
Typical tasks: Optimize decisions, Build an assistant over internal docs, Rank search results, Recommend from what the user just did and 7 more. Companies: Netflix, Yandex, Rabota.ru, Databricks, Sber, Wix and 11 more. Often combined with: Reward model, PPO, Supervised fine-tuning, Multi-armed bandit, Synthetic data generation (LLM) and 66 more.





