Table 1.

Trial and Error Optimization of BPNN with only functional SNPs

HL	U/L	M	Epistasis Models
			1	2	3	4	5
0	0	.1	0.48786	0.49460	0.49560	0.49160	0.49545
0	0	.5	0.48786	0.49460	0.49560	0.49160	0.49545
0	0	.9	0.48786	0.49460	0.49560	0.49160	0.49545
1	5	.1	0.47317	0.45883	0.49568	0.49160	0.49553
1	5	.5	0.36422	0.34229	0.48754	0.49010	0.49543
1	5	.9	0.31206	0.23181	0.34522	0.44670	0.48905
1	10	.1	0.47430	0.46820	0.49607	0.49150	0.49559
1	10	.5	0.35916	0.36446	0.49284	0.49020	0.49542
1	10	.9	0.31209	0.23193	0.34524	0.44660	0.49136
1	15	.1	0.48495	0.47508	0.49599	0.49160	0.49552
1	15	.5	0.37511	0.36221	0.49364	0.49150	0.49542
1	15	.9	0.31217	0.23203	0.34525	0.44670	0.49399
1	20	.1	0.48630	0.49240	0.49583	0.49160	0.49549
1	20	.5	0.40750	0.34406	0.49469	0.49070	0.49544
1	20	.9	0.31217	0.23216	0.34511	0.44660	0.49402
2	5:5	.1	0.49965	0.49997	0.49997	0.50000	0.49996
2	5:5	.5	0.49628	0.49980	0.49996	0.49990	0.49995
2	5:5	.9	0.31205	0.23704	0.41740	0.44670	0.49471
2	10:5	.1	0.49623	0.49980	0.49987	0.49980	0.49972
2	10:5	.5	0.49024	0.49854	0.49929	0.49950	0.49847
2	10:5	.9	0.31201	0.23158	0.35430	0.45450	0.49477
2	15:5	.1	0.49398	0.49944	0.49954	0.49940	0.49913
2	15:5	.5	0.48697	0.49578	0.49850	0.49530	0.49700
2	15:5	.9	0.31199	0.23584	0.35993	0.44740	0.49465
2	20:5	.1	0.49160	0.49849	0.49946	0.49840	0.49889
2	20:5	.5	0.48700	0.49212	0.49808	0.49290	0.49596
2	20:5	.9	0.31199	0.23157	0.34657	0.44750	0.49519

Results from the trial and error optimization of the BPNN on one dataset from each epistasis model. We used 27 different architectures varying in HL – hidden layer, U/L – units per layer, M – momentum. The average classification error across 10 cross-validations from each data set generated for each of the five epistasis models are shown. The best architecture is shown in bold and in Figure 5. This was the most parsimonious architecture with the minimum classification error.