Quantum Enhancements for Deep Reinforcement Learning in Large Spaces

Quantum algorithms have been successfully applied to provide computational speed ups to various machine-learning tasks and methods. A notable exception to this has been deep reinforcement learning (RL). Deep RL combines the power of deep neural networks with reinforcement learning, and has provided some of the most impressive recent artificial-intelligence (AI) results including the famous AlphaGo system—yet, no possibilities for quantum advantages have been identified to date. In this work, we show how quantum computers can enhance the performance of deep RL, especially where the action spaces are large. Specifically, we introduce so-called deep energy-based models, inspired by statistical physics, which we show outperform standard deep RL machinery in learning performance. These models are computationally more demanding, but this can be ameliorated by quantum techniques. Specifically, we provide quantum algorithms, some of which can be run on near-term quantum computers, that can be used to speed up deep energy-based RL. This result opens up a new playing field for quantum enhancements in machine learning and AI.