3. Experiment 2: VQ(A)+VQ(B)+VQ(C).

The sets A, B and C were each modeled with a separate VQ. For set A, 16 codebook vectors were used (cost -116397.7), for B, 19 codebook vectors (cost -3921.6) and for C, 30 codebook vectors (cost -25476.9). The total number of parameters was 1045 and the total cost was -145796.2. The model cost improved compared to Experiment 1 which shows that the general IVGA approach is feasible.

   

Table: Results of the grouping when ${\rm IVGA_{VQ}}$ was run for the combined data set A+B+C. Negative costs are due to leaving out the constant factor in Eq. 1.

Group	Variables	Belong to	Cost	Codebook	Parameters
		data set		vectors
1	1,6	A	-4760.6	7	14
2	2-5,7-8,11	A	-33213.3	10	70
3	9-10,13,15-16,18-21,24-25	A	-50997.9	12	132
4	12,14,17	A	-14201.3	8	24
5	22,26	A	-8233.6	5	10
6	23,27	A	-4920.4	10	20
7	28-34	B	-3889.7	20	140
8	35-38	C	-6678.2	7	28
9	39-42	C	-7248.3	15	60
10	43-44	C	-3140.6	9	18
11	45-46	C	-2708.1	13	26
12	47-50	C	-7215.0	19	76
Total			-147206.8	135	618