MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202611064360 A) filed by Noida Institute Of Engineering And Technology Niet on May 21, 2026, for Symmetry-Aware Graph Neural Network System For Non-Autoregressive Combinatorial Optimization Via Permutation Learning.
Inventors include Dr. Sarika Agarwal; and Rasika Gupta.
The application for the patent was published on July 17, 2026, under issue no. 29/2026.
Abstract: A combinatorial optimization system (100) for solving Travelling Salesman Problem instances on a parallel processing unit (107). The system addresses the technical problem of generating high-quality routing solutions without iterative search or supervised training. The system comprises an equivariant feature extractor (101) that computes covariance-derived canonical coordinate frames with Fourier harmonic encoding to produce rotation-invariant and translation-invariant node features, a scattering-attention graph neural network (102) employing multi-scale diffusion operators with learnable attention-weighted channel mixing, a Gumbel-Sinkhorn permutation operator (103) for differentiable training, and a Hungarian algorithm decoder (104) for discrete permutation extraction. A Monte Carlo dropout controller (105) and snapshot ensemble manager (106) generate diverse candidate solutions from a single trained model. The system achieves improved solution quality relative to classical constructive heuristics through single-pass non- autoregressive inference. The system is applicable to logistics routing, industrial scheduling, and network optimization.
Disclaimer: Curated by HT Syndication.
