A Mutual Information-based Metric for Temporal Expressivity and Trainability Estimation in Quantum Policy Gradient Pipelines

In recent years, various limitations of conventional supervised learning have been highlighted, leading to the emergence of reinforcement learning - and, further, quantum reinforcement learning that exploits quantum resources such as entanglement and superposition - as promising alternatives. Among the various reinforcement learning methodologies, gradient-based approaches, particularly policy gradient methods, are considered to have many benefits. Moreover, in the quantum regime, they also have a profit in that they can be readily implemented through parameterized quantum circuits (PQCs). From the perspective of learning, two indicators can be regarded as most crucial: expressivity and, for gradient-based methods, trainability. While a number of attempts have been made to quantify the expressivity and trainability of PQCs, clear efforts in the context of reinforcement learning have so far been lacking. Therefore, in this study, we newly define the notion of expressivity suited to reinforcement learning and demonstrate that the mutual information between action distribution and reward-signal distribution can, in certain respects, indicate information about both expressivity and trainability. Such research is valuable in that it provides an easy criterion for choosing among various PQCs employed in reinforcement learning, and further, enables the indirect estimation of learning progress even in black-box settings where the agent's achievement aligned with the episodes cannot be explicitly evaluated.