Abstract
The state-of-the-art machine learning approaches are based on classical vonNeumann computing architectures and have been widely used in many industrialand academic domains. With the recent development of quantum computing,researchers and tech-giants have attempted new quantum circuits for machinelearning tasks. However, the existing quantum computing platforms are hard tosimulate classical deep learning models or problems because of theintractability of deep quantum circuits. Thus, it is necessary to designfeasible quantum algorithms for quantum machine learning for noisy intermediatescale quantum (NISQ) devices. This work explores variational quantum circuitsfor deep reinforcement learning. Specifically, we reshape classical deepreinforcement learning algorithms like experience replay and target networkinto a representation of variational quantum circuits. Moreover, we use aquantum information encoding scheme to reduce the number of model parameterscompared to classical neural networks. To the best of our knowledge, this workis the first proof-of-principle demonstration of variational quantum circuitsto approximate the deep $Q$-value function for decision-making andpolicy-selection reinforcement learning with experience replay and targetnetwork. Besides, our variational quantum circuits can be deployed in manynear-term NISQ machines.