9.3 Non-Differentiable Learning of QCBM
The hardware-efficient ansatz we proposed for the QCBM architecture, while simple and intuitive, may be vulnerable to barren plateaus, or regions of exponentially vanishing gradient magnitudes that make training untenable [54, 147, 324]. This provides a strong motivation for exploring a non-differentiable learning alternative such as a Genetic Algorithm (GA).
9.3.1 The principles of Genetic Algorithm
The GA is a powerful evolutionary search heuristic [229] that was introduced in Chapter 3. It performs a multi-directional search by maintaining a population of proposed solutions (chromosomes) for a given problem. Each solution is represented in a fixed alphabet with an established meaning (genes). The population undergoes a simulated evolution, with relatively good solutions producing offspring, which subsequently replace the worse ones, and the quality of a solution is estimated with some objective function (environment...
