Ordering Results

In this experiment we demonstrate EUREKA's ability to pick a method of ordering the tree for expansion. Table 7 shows the results of this experiment. For the fifteen puzzle, the two tested fixed orderings are (Up, Left, Right, Down) and (Down, Left, Right, Up). For the robot arm motion planning domain, two fixed orderings are tested corresponding to ordering joint moves from the base of the arm to the end effector, and ordering joint moves from the end effector first down to the base of the arm last. Only one ordering is used for the artificial domain.

 

Table 7: Ordering Speedup Results

Approach	15Puzzle	Fil-15P	RMP	Fil-15P	Artificial
Ordering 1	49.58	49.58	64.79	58.42	59.22
Ordering 2	52.02	65.41	65.41	80.05	--
TOIDA	50.77	66.13	73.30	79.30	61.03
Local	50.75	68.92	75.37	82.25	60.62
C4.5	50.97	123.79	78.52	87.28	61.03
Combined-C4.5	55.28				--

 

In this experiment, C4.5 yields the best speedup results for all databases, filtered or unfiltered. In the artificial domain, because perfect ordering information is available the TOIDA strategy also yields the best possible speedup results. The combined results are generated using the fifteen puzzle and robot arm motion planning problem instances.

 

Table 8: Ordering Classification Results

Approach	15Puzzle	Fil-15P	RMP	Fil-RMP
Fixed	.6972 (.44)	.7667 (.00)	1.000 (.01)	1.000 (.00)
TOIDA	.6194 (.07)	.7167 (.00)	.3016 (.21)	.3000 (.02)
Local	.6833 (.34)	.5167 (.00)	.6984 (.02)	.7000 (.00)
C4.5	.7069	.1429	.4444	.0000

 

Table 8 shows the results of classifying ordering problems on the filtered and unfiltered data sets. While C4.5 always yields the best average speedup, the learning system does not yield the best classification accuracy on unfiltered data, though it does achieve the best results on the filtered data sets. On the filtered data sets, C4.5 outperforms fixed strategies at a significance value of p$\leq$0.02 or better.