Exponentially Scalable Quantum Reinforcement Learning for Large-Scale Network Scheduling

초록

In modern deep learning research, parameterized quantum circuit (PQC)-based learning algorithms have garnered attention from both industry and academia. Among these, PQC-based reinforcement learning (RL), known as quantum RL (QRL), has garnered particular interest due to its potential for accelerating computations. However, due to the limitations of QRL, the output can only be obtained through quantum measurement computation, which poses scalability challenges. Therefore, this article proposes an exponentially scalable QRL (ES-QRL) to address scalability challenges by deploying basis measurement in environments with extremely large output dimensions, which can be obviously beneficial for large-scale network scheduling. In the literature, conventional QRL algorithms typically rely on classical neural networks (NNs). Thus, they suffer from the need for large-scale qubit utilization to achieve high output dimensions, which is a significant burden in the modern noisy intermediate-scale quantum (NISQ) era. This article leverages our novel measurements, i.e., basis measurement, to overcome the bounds of a limited number of qubits and successfully trains with exponentially large output dimensions with only a few qubits. Our evaluation verifies that the proposed ES-QRL outperforms existing QRL/RL frameworks in large-scale network scheduling problems that require numerous output dimensions. © 2014 IEEE.

키워드