Samyama.ai's graph-grounded optimization beats OR-tools on 7 real-world problems

A new paper from Samyama.ai introduces graph-grounded optimization, a paradigm where real-world optimization problems source decision variables, constraints, and objective coefficients directly from a property knowledge graph via Cypher queries. This replaces traditional free-form text or static tabular inputs. The authors surveyed existing LLM/SLM-driven systems like OptiMUS, Chain-of-Experts, and OPRO, noting that none consume property graphs as primary inputs. They instantiated their approach in the open-source samyama-graph database and evaluated it on seven diverse problems: drug repurposing (245K-node biomedical KG), clinical-trial site selection (7.78M-node registry), Indian supply-chain rerouting (5.34M-node OSM road graph), healthcare equity allocation, economic-environmental grid dispatch, antimicrobial-resistance stewardship (10.4K resistance genes), and wildfire evacuation routing.

They compared a portfolio of Rao-family metaheuristics—BMWR, Jaya, SAMP-Jaya, EHR-Jaya, and Rao-1—against Google OR-tools (CP-SAT and GLOP). Key findings: (i) No single Rao variant dominates across all problem types; BMWR wins on discrete-with-tradeoff and high-dimensional hard-constrained problems, while Rao-1 wins on continuous low-to-mid-dimension problems, supporting a portfolio approach. (ii) OR-tools dominates on small linear/MILP-friendly subproblems but cannot encode the non-linear objectives that emerge in several real-world settings. (iii) Graph-grounded formulations surface data-quality issues—missing properties and degenerate aggregates—that purely text-formulated optimization would silently mask. This work highlights the importance of structured data sources and algorithm diversity for tackling complex optimization tasks.