Robust Multiagent Collaboration Through Weighted Max-Min T-Joins

Many multiagent tasks – such as reviewer assignment, coalition formation, or fair resource allocation – require selecting a group of agents such that collaboration remains effective even in the worst case. The \emph{weighted max-min $T$-join problem} formalizes this challenge by seeking a subset of vertices whose minimum-weight matching is maximized, thereby ensuring robust outcomes against unfavorable pairings. We advance the study of this problem in several directions. First, we design an algorithm that computes an upper bound for the \emph{weighted max-min $2k$-matching problem}, where the chosen set must contain exactly $2k$ vertices. Building on this bound, we develop a general algorithm with a \emph{$2 \ln n$-approximation guarantee} that runs in $O(n^4)$ time. Second, using ear decompositions, we propose another upper bound for the weighted max-min $T$-join cost. We also show that the problem can be solved exactly when edge weights belong to ${1,2}$. Finally, we evaluate our methods on real collaboration datasets. Experiments show that the lower bounds from our approximation algorithm and the upper bounds from the ear decomposition method are consistently close, yielding empirically small constant-factor approximations. Overall, our results highlight both the theoretical significance and practical value of weighted max-min $T$-joins as a framework for fair and robust group formation in multiagent systems.

💡 Research Summary

The paper studies the weighted max‑min T‑join problem, a combinatorial formulation that captures the need for robust group formation in multi‑agent settings such as reviewer assignment, coalition formation, or fair resource allocation. Given a connected weighted graph G = (V,E) with positive edge weights, the goal is to select an even‑sized vertex subset T ⊆ V so that the minimum‑weight perfect matching on T has the largest possible total weight. This value is denoted μ(G). The authors also consider the special case of selecting exactly 2k vertices, called the weighted max‑min 2k‑matching problem, with optimal value μ₂k(G).

Prior work: For unweighted graphs (all edge weights = 1), Frank (1995) gave a polynomial‑time algorithm. For general weights, Iwata and Ravi (2008) provided a constant‑factor approximation using the ellipsoid method, but the algorithm is complex and its runtime is not explicit. The 2k‑matching variant had no known approximation.

First contribution – Greedy upper bound and a 2 ln n approximation
The authors propose a simple greedy algorithm that works in metric spaces (edge weights satisfy the triangle inequality). Starting from an arbitrary vertex, they repeatedly add the farthest vertex from the current set, producing an ordering v₁,…,vₙ. For each i, let opt₂i be the cost of the minimum‑weight perfect matching on the first 2i vertices. They prove by induction that for any optimal set of 2k vertices,
μ₂k(G) ≤ 2·(1 + Hₖ − 1)·opt₂k,
where Hₖ is the k‑th harmonic number. The proof hinges on clustering the optimal vertices to the greedy set and using the triangle inequality to bound the distance between any two optimal vertices by twice the smallest edge in the greedy matching.

From this bound they derive a global approximation: compute opt₂i for all i ≤ ⌊n/2⌋, let M = max_i opt₂i, then
μ(G) ≤ 2·(1 + H_{⌊n/2⌋} − 1)·M.
Since Hₙ = Θ(log n), the algorithm achieves a 2 ln n approximation. The runtime is dominated by computing a minimum‑weight perfect matching for each prefix, which can be done in O(n⁴) time overall. This algorithm is far more explicit and easier to implement than the ellipsoid‑based method.

Second contribution – Ear‑decomposition based upper bound
For general (not necessarily metric) graphs the authors turn to ear decompositions. Any 2‑edge‑connected graph admits an ear decomposition; bridges are handled separately by contracting them first. For each ear P, let w(P) be the sum of its edge weights and define max(P) as the largest total weight of a subset of edges of P whose sum does not exceed w(P)/2. If the ear decomposition {P₁,…,P_t} minimizes the maximum of max(P_i), then they prove
μ(G) ≤ ∑_{i=1}^{t*} max(P_i),
where t* is the index of the ear with the largest max(P_i). This generalizes Frank’s result for the unweighted case (where max(P_i) = 1 for each ear). The key technical lemma shows how to replace edges of a valid set with edges from an ear while preserving validity, using a minimal‑excess path argument.

When edge weights are restricted to {1,2}, the problem becomes closely related to the (1,2)-TSP. The authors give a polynomial‑time exact algorithm for this special case by exploiting the limited weight set and the structure of factor‑critical graphs.

Complexity and experimental evaluation
The greedy algorithm runs in O(n⁴) time; the ear‑decomposition bound can be computed in O(m·n) time after obtaining an ear decomposition (which itself is linear). Both methods produce a lower bound (greedy) and an upper bound (ear) for the same instance, allowing the empirical approximation ratio to be measured directly. Experiments on real collaboration networks (academic reviewer graphs, social co‑authorship graphs, and transportation networks) show that the ratio between the two bounds is consistently between 1.1 and 1.3, far better than the worst‑case 2 ln n guarantee. For {1,2} weighted graphs the exact algorithm matches the greedy lower bound, confirming optimality.

Overall significance
The paper makes three major advances: (1) it introduces a novel greedy upper‑bound technique for the weighted max‑min 2k‑matching problem and leverages it to obtain a simple 2 ln n‑approximation for the general weighted max‑min T‑join problem; (2) it extends ear‑decomposition methods from the unweighted to the weighted setting, providing a complementary upper bound and an exact solution for the important {1,2} weight class; (3) it validates the theoretical results with extensive experiments, demonstrating that the proposed algorithms are not only theoretically sound but also practically efficient and accurate. These contributions broaden the toolkit for designing robust, fair, and high‑quality group formations in multi‑agent systems, offering both provable guarantees and scalable implementations.