In coalition formation with self-interested agents both social welfare of the multi-agent system and stability of individual coalitions must be taken into account. However, in large-scale systems with thousands of agents, finding an optimal solution with respect to both metrics is infeasible. In this paper we propose an approach for finding coalition structures with suboptimal social welfare and coalition stability in large-scale multi-agent systems. Our approach uses multi-agent simulation to model a dynamic coalition formation process. Agents increase coalition stability by deviating from unstable coalitions. Furthermore we present an approach for estimating coalition stability, which alleviates exponential complexity of coalition stability computation. This approach enables us to select a solution with high values of both social welfare and coalition stability. We experimentally show that our approach causes a major increase in coalition stability compared to a baseline social welfare-maximizing algorithm, while maintaining a very small decrease in social welfare.