Training Hybrid Quantum Policies with Classical Surrogates for Actor-Critic Reinforcement Learning
The method introduces a surrogate-assisted actor-critic approach to train hybrid quantum policies for reinforcement learning. It uses a classical neural network, called qtDNN, to locally approximate the quantum circuit during backpropagation. This allows for efficient gradient calculation without differentiating the quantum circuit itself, enabling scalable training on complex tasks like MuJoCo.