Authors:
Saman M. Almufti, Renas Rajab Asaad, Bhopendra Singh, Dinka Lale, Yuri Ryagin, Rahul Panakkal
Addresses:
Department of Information Technology, Technical College of Informatics-Akre, Akre University for Applied Sciences, Akre, Nineveh Governorate, Iraq. Department of Engineering, Amity University Dubai, Dubai, United Arab Emirates. Faculty of Electrical Engineering and Applied Computing, University of Dubrovnik, Dubrovnik, Croatia. Department of Low Temperature Physics and Nanoscale Systems, Ural Federal University, Yekaterinburg, Sverdlovsk Oblast, Russia. Department of Computer Science and Engineering, University of Illinois at Urbana, Champaign, Illinois, United States of America.
The speed reducer benchmark is a mixed-integer nonlinear engineering design problem in which gear geometry, bearing spans, and shaft diameters jointly determine structural weight and eleven stress, deflection, and geometric constraints. This paper develops a continuous U-Turning Ant Colony Optimization (U-TACO) formulation that transfers partial-decision pheromone learning from routing problems to a seven-variable constrained mechanical design. A U-ant samples only a subset of coordinates; the remaining coordinates are completed, the full reducer is decoded and evaluated, and pheromone credit is assigned only to the sampled subset. Complete ants subsequently construct full designs using the same decoder and feasibility-first comparison rule. Eight parameter configurations were evaluated with 50 matched random seeds, while nested samples of 10, 30, and 50 runs were retained for stability analysis. The experiment comprised 400 distinct runs. The precision-search configuration S8 obtained a best feasible weight of 2994.471066, a mean of 2994.471067, a standard deviation of 6.4123×10⁻⁷, and 100% final feasibility. Friedman tests detected configuration effects at each run count; for 50 runs, χ²=259.793658, p=2.277×10⁻⁵², and Kendall’s W=0.742268. Holm-adjusted paired tests favoured S8 over each alternative. The best solution was [3.5, 0.7, 17, 7.3, 7.715320, 3.350215, 5.286654], with four constraints active or numerically near active. The results show that partial-decision reinforcement is effective when combined with exact mixed-integer decoding, constraint-aware completion, and local refinement. Reported runtimes are Python/Numba validation times and are not MATLAB benchmarks.
Keywords: Speed Reducer Design; Gear Geometry; Bearing Spans; Mixed-Integer Decoding; Geometric Constraints; Local Refinement; Comparison Rule; Pheromone Credit.
Received on: 04/07/2025, Revised on: 19/10/2025, Accepted on: 30/10/2025, Published on: 07/10/2026
DOI: 10.69888/FTSSM.2026.000814
FMDB Transactions on Sustainable Structures and Materials, 2026 Vol. 2 No. 2, Pages: 70-87