Learning for Dynamic Combinatorial Optimization without Training Data

We introduce DyCO-GNN, a novel unsupervised learning framework for Dynamic Combinatorial Optimization that requires no training data beyond the problem instance itself. DyCO-GNN leverages structural similarities across time-evolving graph snapshots to accelerate optimization while maintaining solution quality. We evaluate DyCO-GNN on dynamic maximum cut, maximum independent set, and the traveling salesman problem across diverse datasets of varying sizes, demonstrating its superior performance under tight and moderate time budgets. DyCO-GNN consistently outperforms the baseline methods, achieving high-quality solutions up to 3-60x faster, highlighting its practical effectiveness in rapidly evolving resource-constrained settings.

💡 Research Summary

The paper introduces DyCO‑GNN, an unsupervised learning framework for Dynamic Combinatorial Optimization (DCO) that requires no external training data beyond the problem instance itself. Traditional combinatorial optimization (CO) methods either rely on exact solvers, handcrafted heuristics, or supervised machine‑learning approaches that need large labeled datasets and extensive offline training. Recent work on instance‑specific unsupervised learning, notably PI‑GNN, optimizes a graph neural network directly on a single static CO instance by minimizing a differentiable QUBO objective. However, PI‑GNN treats each time‑step of a dynamic problem independently, initializing parameters randomly for every snapshot, which is inefficient when the time between snapshots is short.

The authors first examine a naïve warm‑start baseline that reuses the converged parameters from the previous snapshot as the initialization for the next one. Empirically, warm‑starting accelerates early convergence under tight time budgets, but its advantage quickly disappears as more time is allocated; the model often gets trapped in sub‑optimal local minima because the parameters become over‑confident and gradients vanish.

To overcome this, DyCO‑GNN incorporates a “shrink‑and‑perturb” (SP) strategy originally proposed for supervised learning. When a new graph snapshot G_t arrives, the previous parameters θ_{t‑1} are transformed as
θ_t = λ_shrink·θ_{t‑1} + λ_perturb·ε_t,
where 0 < λ_shrink < 1, 0 < λ_perturb < 1, and ε_t is Gaussian noise (ε_t ∼ N(0,σ²)). Shrinking reduces the magnitude of the weights, thereby lowering the model’s confidence, while the added noise re‑introduces gradient diversity, preventing premature convergence to a poor local optimum. This soft reset preserves useful structural knowledge from the previous snapshot yet encourages exploration of new descent directions.

Algorithmically, DyCO‑GNN proceeds as follows: (1) Randomly initialize θ₁ and fully train PI‑GNN on the first snapshot for a fixed number of epochs (epoch_max). (2) For each subsequent snapshot t = 2…T, apply the SP transformation to obtain θ_t, then perform a limited number of warm‑start epochs (epoch_ws) to fine‑tune the solution for the current QUBO matrix Q_t. The method is applied to three canonical CO problems—Maximum Cut (MaxCut), Maximum Independent Set (MIS), and the Traveling Salesperson Problem (TSP)—all of which can be expressed as QUBO formulations.

Theoretical support is provided via an adaptation of the Goemans‑Williamson (GW) SDP‑based MaxCut algorithm. The authors prove that perturbing the SDP solution before rounding strictly increases the probability of obtaining the optimal cut, establishing that the SP idea has a solid probabilistic justification beyond empirical observation.

Extensive experiments are conducted on five dynamic datasets ranging from a few thousand to tens of thousands of nodes, with 20 snapshots per instance. Three time‑budget regimes are evaluated: urgent (≈0.5 s), moderate (≈2 s), and generous (≈10 s). Performance metrics include Mean Approximation Ratio (Mean ApR) and the number of epochs required to reach a target quality. Results show that DyCO‑GNN consistently outperforms both static PI‑GNN and naïve warm‑started PI‑GNN across all problems and budgets. Specifically, DyCO‑GNN achieves 3–60× faster convergence to comparable or better solution quality, and under moderate to generous budgets it improves Mean ApR by 5–15 % relative to the baselines. Hyper‑parameter sweeps reveal that λ_shrink values around 0.6–0.8 and λ_perturb around 0.1–0.3 yield the best trade‑off between retaining prior knowledge and encouraging exploration.

In summary, DyCO‑GNN demonstrates that (i) instance‑specific unsupervised learning can be extended to dynamic settings without any pre‑collected training data, (ii) a simple shrink‑and‑perturb initialization effectively mitigates the over‑confidence problem of warm‑starting, and (iii) the approach scales to large, real‑world dynamic graphs, making it suitable for time‑critical applications such as network reconfiguration, dynamic routing, and adaptive scheduling. Future work suggested includes handling more complex dynamic constraints (capacity, time windows), extending the framework to non‑graph combinatorial problems, and integrating adaptive perturbation schedules via reinforcement learning.