فهرست نشریات علمی دسترسی آزاد ایران

Hyperparameter optimization (HPO) is essential for maximizing the performance of deep learning models‎. ‎Traditional approaches, such as grid search and Bayesian Optimization (BO), are widely used but can be computationally expensive. ‎ We present Interpolation-Based Optimization (IBO), a novel framework that employs piecewise polynomial interpolation to estimate optimal hyperparameters from sparse evaluations efficiently‎. ‎IBO achieves substantial computational savings by constructing deterministic interpolants with linear per-iteration complexity of O(n.d^3)‎, ‎in contrast to the cubic O(n^3) cost associated with BO‎. ‎Empirical studies on the MNIST dataset show that IBO attains 98.0% accuracy with a 39% reduction in runtime (12 iterations vs. 18) and no statistically significant difference from BO, ‎p = 0.12. In higher-dimensional, lower-cost settings‎, ‎such as ResNet-18 on CIFAR-10‎, performance degrades, highlighting a trade-off between dimensionality and efficiency. ‎More generally‎, ‎IBO is well-suited for resource-constrained settings due to its simplicity, determinism, and computational efficiency. ‎Future work will explore hybrid methods to address scalability problems‎ ‎and extend IBO to more complex modeling architectures‎, ‎such as transformers‎.