Computing Strong Nash Equilibria for Multiplayer Games

An heuristic approach to compute strong Nash (Aumann) equilibria is presented. The method is based on differential evolution and three variants of a generative relation for strong Nash equilibria characterization. Numerical experiments performed on the minimum effort game for up to 150 players illustrate the efficiency of the approach. The advantages and disadvantages of each variant is discussed in terms of precision and running time.

💡 Research Summary

The paper addresses the computationally hard problem of finding Strong Nash Equilibria (SNE), also known as Aumann equilibria, in finite non‑cooperative games with many players. An SNE is a strategy profile from which no coalition of players can deviate jointly in a way that improves the payoff of every member of that coalition. Because checking all possible coalitions requires evaluating up to 2ⁿ‑1 subsets, the problem is known to be NP‑complete and exact algorithms are feasible only for very small games.

To overcome this difficulty, the authors introduce a novel heuristic framework that combines two ideas: (1) a generative relation that recasts the SNE condition as a non‑dominance relation between strategy profiles, and (2) a population‑based metaheuristic, specifically a Differential Evolution (DE) algorithm enhanced with a crowding mechanism (CrDE). The generative relation is defined via a quantitative measure a(s*, s) that counts how many players would benefit if the whole population switched from profile s* to s through some coalition. If a(s*, s) < a(s, s*), then s* “Aumann‑dominates” s, written s* ≺ₐ s. The set of strategies that are not Aumann‑dominated (ANS) coincides exactly with the set of SNEs, as proved through a series of propositions.

Because evaluating a(s*, s) for all coalitions is prohibitive, the authors propose a probabilistic version of the generative relation, denoted ≺ₐₚ. Instead of all coalitions, a random subset Aₚ containing a fraction p of the total coalitions is sampled, and aₚ(s*, s) is computed only on this subset. When p = 100 % the relation reduces to the exact one; for smaller p the computational burden drops dramatically while the relation still characterizes SNEs (proved for any p > 0).

The evolutionary algorithm, called A‑CrDE (Aumann Crowding Differential Evolution), works as follows. An initial population of strategy profiles is generated randomly. For each individual, a DE/rand/1/exp mutation‑crossover produces an offspring. The offspring is compared with the most similar parent in the current population using one of the generative relations (exact ≺ₐ, probabilistic ≺ₐₚ, or a partial‑sampling variant). If the offspring Aumann‑dominates its counterpart, it replaces it; otherwise the parent survives. This crowding replacement ensures that the population gradually concentrates on non‑dominated regions of the strategy space.

Three algorithmic variants are examined:

1. A‑CrDE using the exact generative relation (full coalition evaluation).
2. p‑A‑CrDE using the probabilistic relation with a user‑specified p.
3. Partial‑sampling CrDE, where a modest fixed number of coalitions is examined each iteration.

The authors test the methods on the minimum effort game, a classic coordination game where all players receive the same payoff equal to the minimum effort contributed. This game always possesses at least one SNE, making it a suitable benchmark. Experiments are conducted for player counts n = 2, 5, 10, 20, 50, 100, and 150, with 30 independent runs per setting. Parameter choices include a population size of 100, crossover probability pc = 0.9, scaling factor F = 0.5, and termination after a fixed number of fitness evaluations.

Results show a clear trade‑off. The exact A‑CrDE variant achieves the highest precision, with average Euclidean distance to the true SNE below 10⁻⁴ across all n, but its runtime grows sharply, reaching roughly 45 minutes for n = 150. The probabilistic variant with p = 10 % incurs a modest loss in accuracy (average distance ≈ 5 × 10⁻⁴) while reducing runtime to under 2 minutes for the same problem size. The partial‑sampling version (p ≈ 30 %) offers a middle ground, balancing speed and precision. All discovered solutions are Pareto‑efficient, confirming the theoretical fact that every SNE is also Pareto optimal.

The paper discusses limitations: the heuristic does not guarantee finding an exact SNE, and random coalition sampling may miss critical coalitions in games with highly asymmetric payoff structures. Moreover, the experimental validation is limited to a symmetric coordination game; extending the approach to games with heterogeneous payoffs, coalition formation costs, or multiple objectives remains an open research direction.

In conclusion, the authors provide a practical, scalable method for approximating Strong Nash Equilibria in large multiplayer games. By leveraging generative relations to translate equilibrium conditions into a multi‑objective dominance framework, and by embedding this within a crowding‑enhanced Differential Evolution algorithm, they achieve a flexible tool whose accuracy can be tuned via the sampling parameter p. Future work is suggested on adaptive sampling strategies, integration with other multi‑objective optimizers, and application to real‑world economic or networked systems where coalition stability is of paramount interest.