Bio-Inspired Hyperparameter Optimization for Convolutional Neural Networks: The Artificial Protozoa Optimizer (APO) Approach

Abstract

Hyperparameter configuration remains one of the most consequential and computationally expensive challenges in the design of high-performance Convolutional Neural Networks (CNNs). This paper introduces the Artificial Protozoa Optimizer (APO), a novel bio-inspired metaheuristic algorithm that models the adaptive foraging, dormancy, and reproductive behaviors of protozoan microorganisms to efficiently navigate complex, non-convex hyperparameter search spaces. When integrated with a YOLOv8-based CNN for farmland intrusion detection, APO automated the selection of nine critical hyperparameters—including initial learning rate (lr0), momentum, weight decay, mosaic augmentation, mixup, translation, scale, shear, and horizontal flip probability—yielding transformative and consistent improvements across all performance metrics. Validated on a custom dataset of 1,850 annotated farmland images across three intrusion classes (Human, Animal, No-Intrusion), the APO-CNN system achieved an overall accuracy of 97.84%, a macro-averaged precision of 98.53%, a mean Average Precision at 50% IoU (mAP@0.5) of 98.24%, and a false positive rate of just 0.72%—representing improvements of over 11 percentage points relative to the un-optimized baseline. Benchmarked against 32 state-of-the-art optimization algorithms, APO demonstrated superior convergence behavior and solution quality. The algorithm's parsimonious two-parameter design (neighbor pairs np and maximum proportion fraction pfmax) renders it computationally efficient and practically deployable on resource-constrained IoT edge platforms. This study establishes APO as a compelling and generalizable tool for automated hyperparameter optimization in applied deep learning systems.