Modelling Non-Linear Rendering Processes - Real-Time Rendering: Computer Graphics with Control Engineering - page 56

4.3 EXPERIMENTS
In this section, we describe the experiments conducted to model two real-time ren-
dering applications. The approaches are premised upon the neural network and fuzzy
modelling techniques mentioned in Section 4.2. In all experiments, empirical data
consisting of the per-frame triangle count and frame rate were collected from the two
different applications running on a Pentium IV, 3.2 GHz processor with 2 GB RAM
and NVIDIA's GeForce 6800 graphics board.
In the data collection process, the user is free to move the camera view to simulate
common navigation patterns or object manipulation in virtual environments. This
action is designed so that a wide range of polygon loads and a good combination of
rendering features may be captured. All applications rendered the animated frames
in real time according to the input of the user.
4.3.1 t ime d elay n euRal n etwoRk
To illustrate the applicability of time delay neural networks in modelling the render-
ing process, we selected two applications with different levels of complexity. The
first application was developed to encompass most common rendering parameters in
applications such as textures, fog, lighting, animation, shader effects, and moderate
depth complexity. It consisted of a scene populated by hundreds of instances of a
3D object (a virtual character with a certain surface shading effect) appearing with
an animated landscape. A screenshot of this application is provided in Figure 4.4.
In contrast to the more controlled environment in the first experiment, the appli-
cation in the second experiment was taken from a popular game software system
called “Serious Sam 2”© (2KGames, www.croteam.com). The test case was selected
for its complex rendering functions and scene composition. Figure 4.5 is a screenshot
of this software.
In the second experiment, a certain game environment was selected based on
the level of complexity and the rendering statistics were collected. To capture the
low-level data used in the real-time rendering processes, we used Microsoft's DirectX
tool, PIX Performance Analyzer [44], and utilities from NVIDIA's NVPerfKit [45].
The MATLAB ® Neural Network Plant Identification Tool [46] was utilised for
modelling the rendering process.
In accordance to the system identification methodology described in Chapter 3, a
neural network was first selected as the model structure. The collected data were fed
into the neural network to train it to generate an accurate mapping of the relationship
between the input triangle count and the output frame rate. Different neural network
structures and parameters were tested to determine the best fitting model. This
process continued iteratively until the performance objective (a numeric quantity
describing the difference between the predicted and actual frame rates) was met. The
same procedure was repeated for both experiments.
4.3.2 a daPtiVe n euRo -f uzzy i nfeRence s ystem (anfis)
In addition to neural networks, we introduced the concept of using fuzzy system
modelling for real-time rendering in Section 4.2.2. In Experiment 3, we adopted the
adaptive neuro-fuzzy inference system (ANFIS) to achieve this objective.