CG-CANTS-N: A Versatile Graph-Based Framework for Scalable and Adaptive Problem Solving Across Domains

Neural architecture search (NAS) and neuroevolution have emerged as key methods for designing artificial neural networks (ANNs). While several nature-inspired algorithms, such as Continuous Ant-Based Neural Topology Search (CANTS), have successfully automated the design of recurrent neural networks (RNNs), they suffer from certain limitations, including fixed search constraints and limited exploration strategies. This paper introduces Genetic Programming Collaborative Ant-Based Neural Topology Search (CG-CANTS-N), a novel graph-based NAS framework that employs multiple colonies of simulated ants which move through a continuous search space based on previously placed pheromones. The ant paths through the search space are used to construct graphs which are used as neural architectures. Both the individual ant agents and the ant colonies evolve over time using evolutionary strategies. CG-CANTS-N extends on CANTS by allowing more flexible graph structures, and by utilizing genetic programming functions (e.g., addition, multiplication, trigonometric functions) with trainable weights on graph edges as opposed to traditional neural network neurons. Key innovations include adaptive colony evaporation control, dynamic ant movement strategies, and cycle removal via depth-first search. We demonstrate that CG-CANTS-N is capable of designing graph based genetic programs for time series forecasting tasks which outperform existing state of the art methods.