Routing and Spectrum Allocation (RSA) requires a network controller to select a path for each traffic demand and assign a contiguous block of spectrum that remains consistent across every link in that path. Shared links must also preserve guard bands to prevent interference.
Our research paired a C++ Integer Linear Programming baseline with a generative AI proposal pipeline. The model could suggest paths and spectrum blocks, but every response passed through structured parsing and deterministic feasibility checks before it could be accepted. This kept AI in a narrow, auditable role while code owned constraints, resource accounting, and comparison against an exact optimization model.
- StatusResearch prototype
- Case StudyElastic optical networks
- Exact BaselineIBM ILOG CPLEX
- Core PatternPropose, verify, refine
Routing traffic without breaking spectrum constraints.
Path Selection
Each demand needs a valid source-to-destination path through a directed network without creating, losing, or branching flow.
Spectrum Assignment
Every demand receives one gap-free block of subcarriers, and the same slot indices must continue across every link in its route.
Shared-Link Safety
Allocations using the same link cannot overlap and must preserve the configured guard band between neighboring spectrum blocks.
Exact Baseline
A C++/CPLEX Integer Linear Program models routing, spectrum placement, continuity, guard bands, and shared-link non-overlap, providing ground truth for small and moderate instances.
Giving AI a constrained role inside a verifiable system.
Making proposals machine-checkable and iteratively useful.
Structured Interface
The prompt encodes the network, demands, restrictions, and resource target as JSON. A Python parser validates required fields and normalizes each response into a consistent representation.
Feasibility Checker
Deterministic checks reject invalid paths, nonexistent links, noncontiguous or out-of-range slots, continuity failures, collisions, and insufficient guard bands.
Resource Accounting
The checker independently recomputes data and guard usage, compares it with the declared total, and returns precise failures instead of trusting the model's self-report.
Bounded Refinement
A binary-search controller sets a resource target, verifies the response, tightens the bounds based on feasibility, and retains the best valid candidate.
What I built and what the research demonstrated.
C++ / CPLEX Implementation
I authored the documented RSA implementation, including link-based routing, flow conservation, spectrum placement, path reconstruction, and guarded non-overlap constraints.
Verifier-Controlled AI
As a project co-author, I contributed to a workflow that connected structured prompting with JSON normalization, feasibility checking, and iterative budget control.
Central Finding
Generative AI was most useful as a bounded proposal mechanism. Reliability came from strict schemas, deterministic verification, and an exact optimization reference.
Limits and Next Steps
The model often repeated similar strategies, limiting exploration. Better initial bounds, richer checker feedback, and deliberate strategy diversification could improve the loop.