Distributionally Robust Deep Q-Learning
This research introduces Robust DQN (RDQN), a novel Q-learning algorithm for continuous state spaces with model uncertainty. It uses Distributionally Robust Optimization and Sinkhorn distance to find a worst-case transition, allowing the agent to perform better under model misspecification. RDQN outperforms standard DQN in real-world tasks like portfolio optimization by being more resilient to unfavorable outcomes.