International Artificial Intelligence and Data Processing Symposium (IDAP)
Usage Effect of the 3D Sphere Block in Quantum-Inspired Optimization for Some Benchmark Functions
Büşra Özcan, Abdullah Ateş
Abstract
This paper examines when the three-dimensional Bloch sphere can give better results than the usual two-dimensional model in Quantum-inspired Evolutionary Algorithms. Earlier research has pointed out that the 3D version does better than the 2D one only if the problem has clear links between variables, but there has not been a method to predict this situation. To help with this, a new prediction measure called $\boldsymbol{\psi}$ is introduced. The $\boldsymbol{\psi}$ score brings together three features of the optimization problem. First, it uses the variable interaction ratio, which is based on the off-diagonal elements of the Hessian matrix. Second, it includes the condition number, which shows the difference in curvature in different directions. Third, it considers rotation sensitivity, which is measured by seeing how much the 2D model’s results change if the problem is rotated randomly. Each feature is put on the same scale using the tanh function. A $\boldsymbol{\psi}$ score higher than 0.5 means the 3D model might work better, while a score of 0.5 or less means the 2D model is enough. This method was checked on ten test problems. Three of them had variables that do not affect each other, and seven had variables that do. Each test was repeated thirty times. The results show that the $\boldsymbol{\psi}$ score correctly predicts when the 3D model will be better in seven out of ten problems. In these cases, the 3D model gave better average fitness and more stable results when $\boldsymbol{\psi}$ was 0.5 or higher. However, there were problems with the Rotated Bent Cigar I and II and the Rotated Schwefel II cases. In these cases, $\boldsymbol{\psi}$ was less than 0.5, but the 3D model still did better, which shows that the rotation sensitivity part does not work well for unusual search spaces. The findings suggest that the $\Psi$ score is a good first step, but the type of variable dependency in the problem is more important than just its amount for choosing between 2D and 3D models. In the future, logistic regression will be used to fine-tune the weights for each feature.