Results of reliance

The results in Table 12 show that it is possible to have a form of inference of the strategy used even if the actual $\lambda$ values are not known. It also shows that if the $\lambda$ values are too far from the actual strategies then the brute force styles are more likely to be inferred than the actual strategy. This is most likely due to the amount of information that is present in the brute force techniques, as the quiescent models seem to be more finicky with the weights. The quiescent models are probably skewed by the occasional low level search and thus the information in the higher-level search is made less important. The brute force techniques have clear differences between them as they will always search in a different pattern to that of another brute force strategy, where as the quiescent strategies will sometimes duplicate the search of other quiescent strategies due to non-quiescence.

 

Table: Inference results from inferring search strategies with random $\lambda$ values.

	Actual Strategy
Inferred	min1max1	min1max2	min1max3	min1max4	min2max2
min1max1	100%	28%	59%	38%	10%
min1max2	0%	52%	0%	0%	0%
min1max3	0%	0%	41%	0%	0%
min1max4	0%	0%	0%	43%	0%
min2max2	0%	20%	0%	19%	90%
tolerance $\lambda_{MAT}$	3.992	0.336	0.818	0.142	2.395
tolerance $\lambda_{MOB}$	0.697	0.051	0.025	0.340	0.418
tolerance $\lambda_{CENT}$	0.765	0.421	0.210	0.032	0.459
tolerance $\lambda_{ATT}$	4.414	1.005	0.213	1.500	2.648

From the results in Table 12 it is apparent that the evaluation function does have an impact on the inferred strategy. The most useful element to this analysis is the interval in which each $\lambda$ value was always correct. This tolerance interval shows the maximum distance from the actual $\lambda$ value to the inferred $\lambda$ value that obtains an accurate result. All the results in Table 12 had correct inferences with the $\lambda$ values in-between these domains. Consequently roughly 99% of the time if the inferred $\lambda$ values are within the ranges stated below, then the inferred strategy will be correct. An error of about 0.5 in the $\lambda$ seems to be acceptable in the computing of the search technique. This seems to be a reasonably high amount of leeway and is an encouraging result. This means that errors, which may have occurred in finding the evaluation function, will not affect the inference of the search strategy.