Quantum Multi-Agent Meta Reinforcement Learning

Although quantum supremacy is yet to come, there has recently been an increasing interest in identifying the potential of machine learning (QML) looming era practical computing. Motivated by this, this article we re-design multi-agent reinforcement (MARL) based on unique characteristics neural networks (QNNs) having two separate dimensions trainable parameters: angle parameters affecting output qubit states, and pole associated with measurement basis. Exploiting dyadic trainability as meta-learning capability, propose meta MARL (QM2ARL) that first applies training for meta-QNN learning, followed few-shot or local-QNN training. To avoid overfitting, develop angle-to-pole regularization technique injecting noise into domain during Furthermore, exploiting memory address each trained QNN, introduce concept allowing one save load QNNs using only two-parameter values. We theoretically prove convergence under regularization, simulation corroborate effectiveness QM2ARL achieving high reward fast convergence, well adaptation a time-varying environment.