In terms of quality of the found policy, DICE outperforms JESP for lower horizons: although all methods found (near-)optimal solutions for h = 2,3,4 within the 100 restarts, the variance of JESP is much higher.
Figure 10 shows that while JESP requires much more time than DICE, it does not result in better performance.
Because there are now 3 agents, the number of joint observation histories, and thus the number of entries to compute for exact evaluation and for JESP grows much faster.
In general DICE and DICE-A seem to perform quite well in comparison to JESP. Although JESP'S maximum value is usually equal or greater than the maximum value found by the DICE methods, its mean value is lower and the standard deviation is high.
Again we examined the performance of DICE, DICEA and JESP for this problem, now varying the number of agents.
For h = 3, we see that JESP outperforms the DICE algorithm for all number of agents.
In these hatter experiments we compared against JESP, which, to our knowledge, is the only other approximate planning method for Dec-POMDPs that does not restrict the search space in any way.
On the other hand, because JESP has no parameters, it is somewhat more easy to apply.
However, this work does not intend to present a new state-of-the-art Dec-POMDP solver: we compare against JESP, which is one of the older Dec-POMDP algorithms, and more advanced methods have since been proposed.