Although this work is aimed at giving a general comparison of heuristics for belief space planning, we also present a comparison of the best heuristics within CAltAlt and to some of the other leading approaches to conformant planning. Table 7 lists several features of the evaluated planners, such as their search space, their search direction, whether they are conditional, the type of heuristics, and the implementation language. Note, since each approach uses a different planning representation (BDDs, GraphPlan, etc.), not all of which even use heuristics, it is hard to get a standardized comparison of heuristic effectiveness. Furthermore, not all of the planners use PDDL-like input syntax; MBP, and KACMBP use AR encodings which may give them an advantage in reducing the number of literals and actions. We gave the MBP planners the same grounded and filtered action descriptions that we used in CAltAlt and . We also tried, but do not report results, giving the MBP planners the full set of ground actions without filtering irrelevant actions. It appears that the MBP planners do not use any sort of action pre-processing because performance was much worse with the full grounded set of actions. Nevertheless, Table 8 compares MBP, KACMBP, GPT, CGP, YKA, and CFF with in CAltAlt and in with respect to run time and plan length.
MBP: The MBP planner uses a cardinality heuristic that in many cases overestimates plan distances (as per our implementation with ). MBP uses regression search for conformant plans, but progression search for conditional plans. It is interesting to note that in the more difficult problem instances in the Rovers and Logistics domains MBP and KACMBP tend to generate much longer plans than the other planners. MBP does outperform in some cases but does not find solutions in certain instances (like Rovers 5), most likely because of its heuristic. We note that KACMBP and MBP are quite fast on the Cube Center and Ring domains, but have more trouble on domains like Rovers and Logistics. This illustrates how a heuristic modeling knowledge as opposed to reachability can do well in domains where the challenge is uncertainty not reachability.
Optimal Planners: The optimal approaches (CGP and GPT) tend not to scale as well, despite their good solutions. CGP has trouble constructing its planning graphs as the parallel conformant plan depth increases. CGP spends quite a bit of time computing mutexes, which increases the planning cost as plan lengths increase. CGP does much better on shallow and parallel domains like BT, where it can find one step plans that dunk every package in parallel.
GPT performs progression search that is guided by a heuristic that measures the cost of fully observable plans in state space. GPT finds optimal serial plans but is not as effective when the size of the search space increases. GPT fails to scale with the search space because it becomes difficult to even compute its heuristic (due to a larger state space as well).
YKA: YKA, like CAltAlt is a regression planner, but the search engine is very different and YKA uses a cardinality heuristic. YKA performs well on all the domains because of its search engine based on BDDs. We notice a difference in progression and regression by comparing to YKA, similar to trends found in the comparison between and CAltAlt. Additionally, it seems YKA has a stronger regression search engine than CAltAlt. is able to do better than YKA in the Rovers and Logistics domains, but it is unclear whether that it is because of the search direction or heuristics.
CFF: Conformant FF, a progression planner using a relaxed plan similar to the relaxed plan, does very well in the Rovers and Logistics domains because it uses the highly optimized FF search engine as well as a cheap to compute relaxed plan heuristic. However, CFF does not do as well in the BT, BTC, Cube Center, and Ring problems because there are not as many literals that will be entailed by a belief state. CFF relies on implicitly representing belief states in terms of the literals that are entailed by the belief state, the initial belief state, and the action history. When there are very few literals that can be entailed by the belief state, reasoning about the belief state requires inference about the action history. Another possible reason CFF suffers is our encodings. The Cube Center and Ring domains are naturally expressed with multi-valued state features, and in our transformation to binary state features we describe the values that must hold but also the values that must not hold. This is difficult for CFF because the conditional effect antecedents contain several literals and its heuristic is restricted to considering only one such literal. It may be that CFF is choosing the wrong literal or simply not enough literals to get effective heuristics. However in BT and BTC where we used only one literal in effect antecedents CFF still performs poorly.