Results for generated data

The results of this method, shown in Table 3, are reasonable and again seem to match those which Jansen et al. have established [17][18]. The attribute weights are all approximately what they should be. It can be seen from the amount of compression, that these $\lambda$ values are producing games which are very close to that of the original and the prediction of a game must be very high. The prediction rate will almost never be perfect; this is due to the random probabilistic approach that has been taken. The data is probabilistic, because when given a set of choices, a person will not necessarily make the same choice every time and this is especially true in the game of chess.

 

Table 3: Inference results from self playing using the advanced model.

	No.
Colour	Moves	$\lambda_{MAT}$	$\lambda_{MOB}$	$\lambda_{CENT}$	$\lambda_{ATT}$	I-Random	I-Inferred	Comp (%)
White	74	7.697	0.151	0.499	2.000	120.061	57.205	$47.6\%$
Black	74	7.071	0.339	0.641	3.655	130.271	63.276	$48.6\%$
White	171	9.579	0.179	0.614	3.072	193.409	128.72	$66.6\%$
Black	171	6.905	0.205	0.532	2.810	215.413	132.959	$61.7\%$
White	94	7.924	0.126	0.331	2.954	83.9757	48.662	$57.9\%$
Black	94	6.210	0.218	0.604	2.374	121.842	65.955	$54.1\%$
White	42	6.353	0.216	1.264	2.526	305.108	114.409	$37.5\%$
Black	42	8.001	0.501	1.155	5.370	190.632	22.861	$12.0\%$
White	62	7.170	0.296	-0.915	2.767	107.598	76.778	$48.3\%$
Black	62	5.200	0.084	1.321	3.358	147.366	51.9442	$35.3\%$
W & B	5 Games	6.802	0.199	0.462	2.923	1067.49	1720.6	62.0$\%$
Actual Values		7.0	0.2	0.5	3.0

Notice that larger numbers of moves result in $\lambda$ values that more accurately approximate those of the actual ones. The fact is it can be concluded that more information makes a Maximum Likelihood inference more accurate.