Deep Neural Network-guided PSO for Tracking a Global Optimal Position in Complex Dynamic Environment

Researchers Stephen Raharja and Toshiharu Sugawara have introduced a novel hybrid approach that combines Deep Neural Networks (DNNs) with Particle Swarm Optimization (PSO) to solve a persistent challenge in optimization. Traditional PSO, a heuristic method inspired by bird flocking, struggles in dynamic environments where the optimal solution moves over time. The canonical approach often requires large swarm sizes and complex mechanisms like multiple sub-populations to avoid outdated information and local optima traps. This new research, accepted at the ISMSI 2026 conference, proposes using DNNs to guide the particles, enabling them to learn the characteristics of the environment and adapt their search strategy to predict and pursue a moving target.

The team developed two specific variants: one using a single, centralized neural network for the entire swarm, and another employing distributed networks for each individual particle. In both cases, the DNNs determine particle movement by processing environmental data, allowing the swarm to track shifting global optima more intelligently. Their experimental results, detailed across 10 pages with 5 figures and 5 tables, demonstrate that both AI-enhanced variants achieve a lower cumulative tracking error compared to several recent PSO-based algorithms. Crucially, they accomplish this with a smaller number of particles than there are potential optimal positions, making the process significantly more computationally efficient.

This breakthrough addresses a key limitation in applying optimization algorithms to real-world, time-sensitive problems like robotic path planning in changing terrains, dynamic resource allocation in networks, or real-time financial portfolio optimization. By reducing the required swarm size while improving tracking accuracy, the DNN-PSO hybrid offers a path to faster, more resource-efficient solutions for complex systems that evolve over time.