Conservative Q-Learning (CQL)

Conservative Q-Learning (CQL)

Offline reinforcement learning method that actively penalizes overestimated Q-values to keep the policy close to the behavioral data distribution and prevent divergence.

← Geri