Ito Brownian - Continuous Ant Colony Optimization (IB-CACO)

Abstract

We advance the study of optimization over loss landscape surfaces (LLS), with particular emphasis on the highdimensional
and nonconvex settings characteristic of modern machine learning (ML). Building on our earlier work that introduced the Brownian Bridge–Continuous Ant Colony Optimization (BB-CACO) algorithm, we extend the approach to a more general framework based on Itˆo bridges. The resulting algorithm, Itˆo Bridge–CACO (IB-CACO), dynamically adjusts drift and diffusion parameters to enhance exploration and convergence. Unlike stochastic gradient descent (SGD) and its common variants (e.g., RMSProp, Adam), which are often hindered by noise, sensitivity to hyperparameters, and entrapment in local minima, IB-CACO consistently demonstrates superior performance in both locating global minima and achieving computational efficiency. Beyond ML benchmarks and training of Large language models (LLMs), the algorithm shows promise in domains where rapid and reliable global optimization is critical, including aircraft design, natural resource prospecting, drone navigation, and autonomous rescue operations.