I've have recast the problem of partner selection in n-player games, with stochastic strategies, as a problem of selecting the right combination of players. To support this approach, each player maintains a pool of partner combinations and a probability it associates to each combination. After playing with a combination if the outcome was considered acceptable, nothing happens. Otherwise, the probability of the combination is decreased and distributed among the others and a new combination is selected. The end result is that the good combinations once in the pool never leave and absorb the decreasing probabilities of bad combinations.
I've developed two demonstrators of the partner selection model.