First you'd have a heuristic that ranks outcomes from turn 3, which would just mean weighing the pieces the computer lost against the pieces the opponent lost. Then you rank each turn-2 node based on how desirable the turn-3 leaf-nodes it leads to are and how likely the opponent is to actually make the moves that lead to it. Then you'd do the same for turn-1 and then pick the move that scored highest.