Memetic Metaheuristic Initialisation of Neural Network Surrogates for Weight Optimisation of Steel Roof Trusses

GA- and PSO-initialised MLP surrogates were compared with a plain MLP for steel roof trusses under an equal training budget and a validation-based memetic fitness. The main conclusions are: 1.      Both hybrids improved on the base MLP on all four profiles. MLP-GA gained 0.005–0.012 in mean testing R² and won 27 of 32 profile–response combinations; MLP-PSO gained 0.004–0.008 and won 20. 2.      The memetic fitness is the setting that makes hybridisation reliable. Scoring untrained weights gave small and inconsistent gains. 3.      The gain concentrates on the hardest response, maximum utilisation, and is small for smooth responses such as weight. Hybridisation is worthwhile mainly where the base network struggles. 4.      Used as a design tool at 24 spans, all 72 designs selected by the perceptron surrogates complied with IS 800:2007. MLP-GA gave up 2.4 % of weight on average against the true optimum, compared with 4.3 % for the base MLP. 5.      Optimising against a prediction biases the selected design towards under-predicted weight and utilisation. Every design proposed by a surrogate should be confirmed by full analysis before use. Future work should extend the database to designs with spare capacity so that an acceptance margin can operate, and verify designs proposed from the full section catalogue with independent analyses.

1.      Cho, Y., Xia, L., Hong, S., Kim, S. and Bae, J. 2007. Study of optimized steel truss design using neural network to resist lateral loads. Key Engineering Materials, 348–349: 405–408. 2.      Clerc, M. and Kennedy, J. 2002. The particle swarm – explosion, stability, and convergence in a multidimensional complex space. IEEE Transactions on Evolutionary Computation, 6(1): 58–73. 3.      IS 800:2007. General Construction in Steel – Code of Practice. Bureau of Indian Standards, New Delhi. 4.      IS 875:1987. Code of Practice for Design Loads (Other than Earthquake) for Buildings and Structures, Parts 1–3. Bureau of Indian Standards, New Delhi. 5.      Kao, C. Y., Hung, S. L. and Xu, P. J. 2025. Application of the artificial neural network and enhanced particle swarm optimization to model updating of structures. Journal of Civil Structural Health Monitoring, 15: 59–86. 6.      Kashyap, G. S., Nafis, M. T. and Wazir, S. 2025. A study of hybrid and evolutionary metaheuristics for single hidden layer feedforward neural network architecture. arXiv. 7.      Kennedy, J. and Eberhart, R. 1995. Particle swarm optimization. In: Proceedings of ICNN’95 – International Conference on Neural Networks, 4: 1942–1948. 8.      Kulyal, D. 2026. Development of a Hybrid Artificial Neural Network and Metaheuristic Optimization Framework for the Optimal Design of Steel Trusses. Ph.D. thesis, G.B. Pant University of Agriculture & Technology, Pantnagar, India. 9.      Kulyal, D. and Verma, V. K. Under review. Additive versus multiplicative neuron surrogates for predicting the structural response of steel roof trusses. 10.   Lee, D., Noh, S., Kim, J. and Lee, S. 2025. Efficient hyperparameter optimization using metaheuristics for machine learning in truss steel structure cross-section prediction. Buildings, 15(8): 2791. 11.   Moscato, P. 1989. On Evolution, Search, Optimization, Genetic Algorithms and Martial Arts: Towards Memetic Algorithms. Caltech Concurrent Computation Program, C3P Report 826. 12.   Nourian, N., El-Badry, M. and Jamshidi, M. 2023. Design optimization of truss structures using a graph neural network-based surrogate model. Algorithms, 16(8): 380. 13.   Ojha, V. K., Abraham, A. and Snášel, V. 2017. Metaheuristic design of feedforward neural networks: A review of two decades of research. Engineering Applications of Artificial Intelligence, 60: 97–116.