Abstract
Wildlife conservation applications demand neural network architectures that simultaneously optimize prediction accuracy, computational efficiency, and model interpretability-a challenge inadequately addressed by existing single-objective methods. We present a novel multi-objective hybrid algorithm combining genetic algorithms, simulated annealing, and reinforcement learning for conservation-specific neural network optimization. Our approach uniquely formulates conservation objectives through species identification accuracy, habitat modeling precision, and real-time deployment constraints while maintaining model transparency for conservation practitioners. The algorithm integrates adaptive temperature scheduling responsive to population diversity and a conservation-aware reward function incorporating ecological domain knowledge. Theoretical analysis establishes convergence guarantees under conservation-specific constraints. Comprehensive evaluation on established wildlife datasets demonstrates 34% improvement in hypervolume indicator and 42% reduction in computational overhead compared to state-of-the-art multi-objective algorithms including NSGA-III, RVEA, MOEA/DD, and recent transformer-based methods. The framework successfully balances multiple competing objectives while providing interpretable solutions for conservation decision-making, advancing automated neural architecture search for ecological applications with immediate practical applicability.