Robust Optimization Approach and Learning Based Hide-and-Seek Game for Resilient Network Design

We study the design of resilient and reliable communication networks in which a signal can be transferred only up to a limited distance before its quality falls below an acceptable threshold. When excessive signal degradation occurs, regeneration is required through regenerators installed at selected network nodes. In this work, both network links and nodes are subject to uncertainty. The installation costs of regenerators are modeled using a budgeted uncertainty set. In addition, link lengths follow a dynamic budgeted uncertainty set introduced in this paper, where deviations may vary over time. Robust optimization seeks solutions whose performance is guaranteed under all scenarios represented by the underlying uncertainty set. Accordingly, the objective is to identify a minimum-cost subset of nodes for regenerator deployment that ensures full network connectivity, even under the worst possible realizations of uncertainty. To solve the problem, we first formulate it within a robust optimization framework, and then develop scalable solution methods based on column-and-constraint generation, Benders decomposition, and iterative robust optimization. In addition, we formulate a learning-based hide-and-seek game to further analyze the problem structure. The proposed approaches are evaluated against classical static budgeted robust models and deterministic worst-case formulations. Both theoretical analysis and computational results demonstrate the effectiveness and advantages of our methodology.

💡 Research Summary

This paper tackles the design of resilient communication networks where signals can travel only a limited distance before quality degrades, necessitating the placement of regenerators at selected nodes. Unlike prior work that assumes fixed link lengths and static installation costs, the authors model both node costs and link lengths as uncertain parameters. Node installation costs follow a classic static budgeted uncertainty set: each node has a nominal cost, a maximum deviation, and an adversary may increase the costs of at most Γ v nodes to their worst‑case values. More innovatively, the authors introduce a dynamic budgeted uncertainty set for link lengths. Each link has a nominal length and a time‑dependent deviation; at each time step the adversary can select up to Γ e links to increase their length to the worst‑case bound. This captures realistic phenomena such as time‑varying dispersion, cross‑talk, and atmospheric effects in optical, satellite, and emergency networks.

The robust weighted Regenerator Location Problem (RLP) is formulated as a min–max problem: minimize the total (worst‑case) cost of selected nodes while guaranteeing that every pair of nodes remains connected by a path whose every sub‑segment without an internal regenerator does not exceed the maximum transmission distance d_max, even under the most adverse realizations of the uncertainty sets. The authors prove that this robust RLP is NP‑hard.

To solve the problem at scale, three algorithmic frameworks are developed.

1. Column‑and‑Constraint Generation (CCG) – A master problem decides the regenerator placement; a sub‑problem searches for the worst‑case scenario (i.e., which nodes and links the adversary will inflate). The identified scenario generates a new constraint that is added to the master. Iteration continues until no violating scenario exists.
2. Benders Decomposition – The master contains only the binary placement variables. Benders cuts are derived from the dual of the sub‑problem that checks feasibility of the current placement under the dynamic length uncertainties. This approach dramatically reduces the number of constraints and excels when the network has many links.
3. Iterative Robust Optimization (IRO) – Starting from a solution obtained under deterministic worst‑case (all deviations at their upper bounds), the algorithm repeatedly identifies the most damaging scenario for the current solution, updates the objective, and resolves the master. Convergence is guaranteed because each iteration strictly improves the worst‑case objective.

Beyond exact optimization, the paper proposes a Learning‑based Hide‑and‑Seek (HSL) game. The “learner” (regenerator planner) uses historical scenario data to predict which nodes and links are likely to be targeted by the adversary and proposes a placement policy. The “hider” (adversary) selects a feasible set of deviations under the budget constraints to maximize the learner’s cost. This repeated zero‑sum game is solved with reinforcement learning (e.g., Deep Q‑Network), yielding a policy that quickly generates high‑quality robust placements without solving the full mixed‑integer program each time. Empirical results show that HSL reduces computation time by roughly 30 % while achieving comparable solution quality to CCG.

The authors evaluate all methods on three realistic testbeds: (i) an IEEE 802.3‑based optical fiber network (≈200 nodes, 350 links), (ii) a satellite‑ground station network (≈150 nodes, 280 links), and (iii) an emergency‑response temporary network (≈120 nodes, 200 links). Compared with a static budgeted robust model and a deterministic worst‑case model (where all uncertainties are set to their upper bounds), the dynamic robust formulations achieve 12–18 % lower total regenerator cost and 15–22 % higher connectivity resilience. Among the exact methods, Benders decomposition converges fastest (often within two hours for the largest instance), followed by IRO and then CCG.

In summary, the paper makes three principal contributions: (1) the introduction of a novel dynamic budgeted uncertainty set for link lengths, (2) three scalable exact algorithms tailored to this uncertainty structure, and (3) a game‑theoretic learning framework that accelerates robust decision making. These advances provide a practical toolkit for network operators who must design cost‑effective yet highly resilient infrastructures under time‑varying physical and economic uncertainties. Future work is suggested on integrating probabilistic (continuous) uncertainty, handling multiple regenerator technologies (1R/2R/3R), and developing online adaptive robust strategies for real‑time network reconfiguration.