An Anytime Algorithm for Optimal Coalition Structure Generation

Coalition formation is a fundamental type of interaction that involves the creation of coherent groupings of distinct, autonomous, agents in order to efficiently achieve their individual or collective goals. Forming effective coalitions is a major research challenge in the field of multi-agent systems. Central to this endeavour is the problem of determining which of the many possible coalitions to form in order to achieve some goal. This usually requires calculating a value for every possible coalition, known as the coalition value, which indicates how beneficial that coalition would be if it was formed. Once these values are calculated, the agents usually need to find a combination of coalitions, in which every agent belongs to exactly one coalition, and by which the overall outcome of the system is maximized. However, this coalition structure generation problem is extremely challenging due to the number of possible solutions that need to be examined, which grows exponentially with the number of agents involved. To date, therefore, many algorithms have been proposed to solve this problem using different techniques ranging from dynamic programming, to integer programming, to stochastic search all of which suffer from major limitations relating to execution time, solution quality, and memory requirements. With this in mind, we develop an anytime algorithm to solve the coalition structure generation problem. Specifically, the algorithm uses a novel representation of the search space, which partitions the space of possible solutions into sub-spaces such that it is possible to compute upper and lower bounds on the values of the best coalition structures in them. These bounds are then used to identify the sub-spaces that have no potential of containing the optimal solution so that they can be pruned. The algorithm, then, searches through the remaining sub-spaces very efficiently using a branch-and-bound technique to avoid examining all the solutions within the searched subspace(s). In this setting, we prove that our algorithm enumerates all coalition structures efficiently by avoiding redundant and invalid solutions automatically. Moreover, in order to effectively test our algorithm we develop a new type of input distribution which allows us to generate more reliable benchmarks compared to the input distributions previously used in the field. Given this new distribution, we show that for 27 agents our algorithm is able to find solutions that are optimal in 0.175% of the time required by the fastest available algorithm in the literature. The algorithm is anytime, and if interrupted before it would have normally terminated, it can still provide a solution that is guaranteed to be within a bound from the optimal one. Moreover, the guarantees we provide on the quality of the solution are significantly better than those provided by the previous state of the art algorithms designed for this purpose. For example, for the worst case distribution given 25 agents, our algorithm is able to find a 90% efficient solution in around 10% of time it takes to find the optimal solution.

💡 Research Summary

Coalition Structure Generation (CSG) is a cornerstone problem in multi‑agent systems: given a set of autonomous agents, one must partition them into disjoint coalitions so that the sum of the coalition values is maximized. The number of possible partitions grows as the Bell number Bₙ, making exhaustive search infeasible even for modest n (e.g., B₂₅ ≈ 4.6 × 10¹⁶). Existing approaches fall into three broad categories. Dynamic‑programming (DP) methods guarantee optimality but require O(3ⁿ) memory, which quickly becomes prohibitive. Integer‑programming (IP) formulations can be solved with commercial solvers, yet they often fail to converge within reasonable time for larger instances. Stochastic or meta‑heuristic techniques (genetic algorithms, simulated annealing, MCMC) are fast but provide no deterministic quality guarantees and lack anytime behavior.

The authors propose a novel anytime algorithm that overcomes these limitations by restructuring the search space and exploiting tight upper‑ and lower‑bounds for systematic pruning. The key contributions are:

1. Search‑space representation – The set of all coalition structures is encoded as a binary tree where each node corresponds to a partial partition. This representation eliminates duplicate generation automatically because each complete partition can be reached by a unique path.

2. Sub‑space decomposition and bounding – At a chosen depth of the tree, all continuations sharing the same prefix form a “sub‑space”. For each sub‑space the algorithm computes:

   • Upper bound – the maximum possible value that any coalition structure within the sub‑space can achieve. This is derived by taking the highest coalition value among the remaining agents and multiplying it by the number of agents still to be assigned, yielding a conservative but often tight estimate.
   • Lower bound – the value of the best complete coalition structure discovered so far.

3. Branch‑and‑bound pruning – If a sub‑space’s upper bound is lower than the current global lower bound, the sub‑space cannot contain the optimal solution and is discarded without further exploration. Otherwise the algorithm proceeds with a depth‑first or best‑first (A*) traversal, updating the lower bound whenever a better structure is found, which in turn may prune additional sub‑spaces.

4. Anytime property – The algorithm can be interrupted at any moment. The incumbent lower bound and its associated coalition structure are returned, together with a provable guarantee: the optimal value lies between the incumbent lower bound and the maximum upper bound among all remaining sub‑spaces. Consequently, even a prematurely stopped run yields a solution with a known worst‑case deviation from optimality.

5. New benchmark distribution – Recognizing that the commonly used Uniform, Normal, and Scaled‑Normal distributions do not capture realistic inter‑coalition correlations, the authors design a “correlated scaled” distribution. Coalition values increase non‑linearly with coalition size, and larger coalitions exhibit higher variance, mimicking many real‑world scenarios (e.g., economies of scale with uncertain returns). This distribution provides a more stringent and representative testbed for CSG algorithms.

Experimental evaluation focuses on instances with 25–27 agents, comparing the proposed method against the state‑of‑the‑art DP‑based algorithm (IDP), an integer‑programming approach, and a leading stochastic method (ODP). Results show dramatic improvements: for 27 agents the new algorithm reaches the optimal coalition structure in only 0.175 % of the time required by the fastest existing method. For the worst‑case distribution with 25 agents, a solution that is at least 90 % of optimal is obtained after roughly 10 % of the time needed to compute the exact optimum. Memory consumption is reduced by more than 80 % relative to DP because only one upper and one lower bound per sub‑space need to be stored (O(2ⁿ) versus O(3ⁿ)).

Theoretical analysis confirms that the number of sub‑spaces is bounded by a subset of the Bell number, giving a worst‑case time complexity of O(2ⁿ) for bound calculations. In practice, the bound‑driven pruning eliminates the majority of sub‑spaces, yielding an average‑case complexity that behaves almost polynomially for the tested sizes. The algorithm’s correctness proof demonstrates that every feasible coalition structure is either examined directly or implicitly pruned by a provably tighter bound, guaranteeing that the final solution is optimal if the search runs to completion.

In summary, this paper introduces a powerful anytime framework for coalition structure generation that simultaneously addresses three critical challenges: exponential search space, prohibitive memory requirements, and lack of quality guarantees in existing methods. By partitioning the search space, computing tight bounds, and pruning aggressively, the algorithm achieves orders‑of‑magnitude speedups while delivering provable solution quality at any interruption point. The new benchmark distribution further strengthens the empirical validation, making the results more credible for real‑world applications. Future work may extend the approach to dynamic environments (agents entering or leaving), non‑additive coalition value functions, and distributed implementations where each agent contributes to bound computation, opening promising avenues for scalable multi‑agent coordination.