Graphics Reference
In-Depth Information
Measured Output vs User Reference Levels
405
Reference
Measured output
400
395
390
385
380
375
370
365
360
355
0
2
4
6
8
10
12
14
16
×10 4
Frame
FIGURE 4.11
Measured and reference output from ANFIS in Experiment 3.
experiment was less than 5 FPS. A delay registered in the experiment results may be
attributed to the latency between the connected peripherals in the experiment setup.
4.5 DISCUSSION
The common objective of our tests was to derive an accurate system model of the
rendering processes. While it may seem ideal to have a single model for all appli-
cations, a single model is impractical because various rendering applications have
different dynamics and vary in the numbers of components contributing to the final
render time.
For example, applications differ in the types and numbers of processes such as net-
work communication, application logic, and input-output computations. Hence it is not
a trivial task to derive a universal model for all rendering applications. Furthermore, a
generalised model would not necessarily be useful because it might not provide a user
with a set of components that could be used easily in the rendering process.
Another benefit from using soft computing techniques such as neural networks
and fuzzy systems is that they provide greater convenience for modelling wider
operating ranges compared to using linear model structures. They eliminate the
need to conduct several tedious data collection procedures over an operating range.
Furthermore, when a satisfactory model is derived, there is no need to re-train the
neural network or ANFIS unless the construct of the application changes. As to speed
of modelling, the training of our neural networks typically required fewer than 3 min-
utes for a dataset of approximately 5,000 data points on a mid-end desktop computer.
Search WWH ::




Custom Search