Q-Learning in enormous action spaces via amortized approximate maximization
This addresses the problem of scalability in reinforcement learning for researchers and practitioners dealing with large action spaces, though it is an incremental improvement over existing Q-learning methods.
The paper tackles the challenge of applying Q-learning to high-dimensional or continuous action spaces by replacing the expensive maximization over all actions with a maximization over a small subset sampled from a learned proposal distribution, resulting in Amortized Q-learning (AQL) that outperforms D3PG and QT-Opt on continuous control tasks with up to 21-dimensional actions and efficiently handles spaces with thousands of discrete actions.
Applying Q-learning to high-dimensional or continuous action spaces can be difficult due to the required maximization over the set of possible actions. Motivated by techniques from amortized inference, we replace the expensive maximization over all actions with a maximization over a small subset of possible actions sampled from a learned proposal distribution. The resulting approach, which we dub Amortized Q-learning (AQL), is able to handle discrete, continuous, or hybrid action spaces while maintaining the benefits of Q-learning. Our experiments on continuous control tasks with up to 21 dimensional actions show that AQL outperforms D3PG (Barth-Maron et al, 2018) and QT-Opt (Kalashnikov et al, 2018). Experiments on structured discrete action spaces demonstrate that AQL can efficiently learn good policies in spaces with thousands of discrete actions.