Stacking generalization models for classification are ensemble techniques that use a meta-learner to combine the results of several base learners. Most often results reported by base learners vary significantly, so a trade-off is desired from the meta-learner. This paper proposes a game theoretic approach that uses the Nash equilibrium concept for building the meta-learner. The equilibrium of the game is approximated using a differential evolution algorithm. Numerical experiments are used to illustrate the potential of the approach.