Event-Based Deep Reinforcement Learning for Quantum Control
Haixu Yu, Xudong Zhao · IEEE Transactions on Emerging Topics in Computational Intelligence · 2023
Deep reinforcement learning (DRL) has been widely used in many quantum control tasks, in which rewards are used to guide the agent in acquiring an optimal control strategy. However, designing suitable rewards is usually challenging, which seriously affects the DRL's application performance, especially for complex quantum control tasks. In this article, we propose a simple and general reward design method by constructing some events, where events are defined by fidelity intervals. In the proposed event-based DRL (EDRL), events are used to describe the evolution of a quantum system. Rewards are designed according to several defined events rather than a large number of specific evolved quantum states, thereby resulting in a reduction in dimensions to be considered. With designed event-attributed rewards, the agent can capture the main evolution characteristics of a quantum system, and then obtain an effective control strategy. The effectiveness of the proposed EDRL method is verified by three quantum systems: i.e., 1) the one-qubit closed quantum system, 2) the two-qubit closed quantum system, and 3) the two-level open quantum system. Numerical comparisons demonstrate that the proposed EDRL outperforms the traditional DRL algorithms (deep Q-network and proximal policy optimization) with some representative reward methods for solving quantum control problems.