of IDA*IM begun to significantly decrease just before 10 iterations. When the histogram size was increased to 100 the prediction accuracy begun to decrease at around 20 iterations. When we increased the histogram size to 500, the drop off in accuracy went away and the growth factor remained around the correct value of 2 until the very final iterations of search. Finally, we increased the histogram size much further to 10,000, the performance was very similar to the performance of 500, including the slight estimation inaccuracy during the final few iterations.Figures 10b and 10c compare IDA*, IDA*CR , and IDA*IM with a histogram size of 500 on the pancake problem. As before we can see that IDA* performs poorly as it requires many iterations, each of which is very similar to the previous. IDA*CR , once again grows its iterations quite quickly, about quadrupling each subsequent iteration. The incremental model, using sufficiently large histograms, is rather accurate at doubling the size of its iterations, however, its performance is similar to IDA*CR with respect to the number of instances solved in the time limit. We suspect that IDA*IM does not outperform IDA*CR in this domain due to its increased overhead in maintaining larger histograms.5.4 SummaryWhen trained off-line, the incremental model was able to make predictions on the 15-puzzle domain that were nearly indistinguishable from CDP, the current state-of-the art. In addition, the incremental model was able to estimate the number of node expansions on a real-valued variant of the sliding tiles puzzle where each move costs the square root of the tile number being moved. When presented with pairs of 15-puzzle instances, the incremental model trained with10 samples was more accurately able to predict which instance would requirefewer expansions than CDP when trained with 10,000 samples.The incremental model made very accurate predictions across all domains when trained on-line and when used to control the bounds for IDA*, our model made for a robust search. While each of the alternative approaches occasionally gave extremely poor performance, IDA* controlled by the incremental model achieved the best performance of the IDA* searches in the vacuum maze and uniform tree domains and was competitive with the best search algorithms for both of the sliding tiles domains and the pancake puzzle. This provides an ex- ample of how a flexible tree model can be used in practice.6 DiscussionIn search spaces with small branching factors such as the vacuum maze domain, the back-off model seems to have a greater impact on the accuracy of predic- tions than in search spaces with larger branching factor such as the sliding tiles domains. Because the branching factor in the vacuum maze domain is small, however, the simulation must extrapolate out to great depths (many of which the model has not been trained on) to accumulate the desired number of ex- pansions. The simple back-off model used here merely ignored depth. While this
đang được dịch, vui lòng đợi..
