In this online appendix, we analyse the number of feasible
coalitions given different constraints, evaluate our Divide and Conquer (D&C) algorithm, and then evaluate our CSG algorithm.
Analysing the Number of Feasible Coalitions
Figure 1 shows the ratio between the number of all coali- tions and the number of feasible ones. More specifically, given different numbers of agents, Figure 1(A) shows the case where the constraints are all positive (i.e., N = ∅), while Figure 1(B) shows the case where they are all neg- ative (i.e., P = {∅}). As can be seen, the percentage of feasible coalitions increases significantly given an increase in the number of positive constraints, or given a drop in the number of negative ones. To analyse this effect even further, we show in Figure 1(C) the case where there are both positive and negative constraints, with equal number (e.g., given 2000 constraints, a 1000 would be positive, and the other 1000 would be negative). In particular, as the number of constraints increases, the influence that the posi- tive constraints have on the percentage of feasible coalitions becomes smaller, while the influence of the negative con- straints becomes larger. Therefore, as the number of con- straints increases, the percentage of feasible coalitions first increases and then gradually starts to decrease. Moreover, the more agents we have, the fewer constraints are needed before this decrease starts to occur.
Evaluating our Divide and Conquer Algorithm
Given different numbers of agents, Figure 2 shows, on a log scale, the ratio between the time taken by our Divide and Conquer (D&C) algorithm, and that taken a Brute Force (BF) algorithm, to generate the feasible coalitions. In more detail, D&C first generates the transformed set of constraints T ∗ , and then generates the coalitions that correspond to ev- ery (P ∗ , N ∗ ) in T ∗ . On the other hand, the BF algorithm checks every coalition against every constraint in order to verify whether it is a feasible one. As can be seen in the figure, D&C is faster by orders of magnitude. Figure 1: The ratio between the number of all coalitions and Evaluating our CSG algorithm for CCF the number of feasible ones. In Figures (A), the constraints Given different numbers of agents, Figure 3 presents the ter- are only positive. In (B), the constraints are only negative. mination time of our CSG algorithm for 1000 positive and In (C), the constraints are half positive and half negative. 1000 negative constraints in seconds (logscale). It is in- teresting to observe similar curved pattern that follows the number of coalitions that can be seen in Figure 1(C). Finally, we evaluate the anytime property of our algo- rithm, and that is by computing the ratio between the value of the optimal solution (which was obtained at the end of the run), and the best solution that the algorithm found at any point in time during that run. Figure 4 shows this evaluation given different numbers of agents and constraints. As can be seen, in most of the cases, the value of our initial solution is more than 90% of the optimal one. It can also be seen that the algorithm finds an optimal solution very quickly af- ter around 20% of its running time, and the rest of the time is spent to confirm that no better solution exists.
Figure 2: The figure shows, on a log scale, the ratio be-
tween the time to generate the feasible coalitions using the Brute Force (BF) algorithm and using our Divide and Con- quer (D&C) algorithm.
Figure 3: The running time of our CSG algorithm in sec (log
scale).
Figure 4: How the solution quality grows over the running