Biomedical Engineering Reference
In-Depth Information
(Building block test vs. time) one algorithm and multiple farms
6,000
1Alg0Farms
1Alg2Farms
1Alg4Farms
1Alg8Farms
ANN
5,000
4,000
3,000
2,000
1,000
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Building block test
Figure 19.2
This figure illustrates the achievable 16.7 speedup, which may be realized when
implementing a 1K neuron ANN within the fmGA search process. This speedup is not direct and
final because there is still runtime required for the local search, which must be accomplished
on good solutions found by the ANN.
increases, the average execution time decreases for any given BB test. In this test,
there exists a significant improvement in modifying the number of farming nodes
from 0 to 2 and from 2 to 4. An increase in the farming nodes from 4 to 8 provides a
small improvement. The best speedup obtained was with 8 farming nodes where the
serial time was 5080 s, whereas the parallel time was 1684 s yielding a speedup of
three times. This validates our model and we can draw a conclusion that this model
increases the efficiency.
Efficiency measurements were taken from run times of the original fmGA using the
CHARMm code versus a mach-up neural network code in place of the CHARMm
function. Table 19.6 summarizes the results. The ANNs are shown to provide an
increase in efficiency over the CHARMm code. Figure 19.3
a
graphically illustrates
this speedup. Notice that the 25 neuron RBFNN is not listed. It can be approximated to
take 0.021 ms, which is easily more efficient than the CHARMm code. Additionally,
Figure 19.3
b
is provided to illustrate the increase in calculation time spent as the
number of neurons is added to the neural network. Furthermore, the ANN also can
yield a preprocessing speedup of 16.7 over the farming model parallel implementation.
This speedup is illustrated in Figure 19.2, where the ANN is shown to be more
efficient than every farming implementation tested.
TABLE 19.6 Average Time for Evaluations on PIII 800 MHz
(Efficiency Results)
Application
msec
CHARMm MET(POLY)
6.211(41.04)
One nonlinear layer having
(0.007,0.008,0.012,0.068,0.640)
(3,4,10,100,1K) neurons
Search WWH ::
Custom Search