This thesis describes implementations of motion control systems that are based on fuzzy logic; fuzzy
motion controllers. The controllers are used by to drive a variety of simulated vehicles and computeranimated
characters. The problem of heading towards a destination whilst simultaneously avoiding
static and dynamic obstacles is addressed with fuzzy motion controllers. For situations where a level
above reactive motion control is required, such as path-planning behaviour or traffic rule following,
then hybrid algorithms are proposed; combining fuzzy motion controllers with other navigation
algorithms. Consideration is given to behavioural level of detail models, with transition between behavioural
models of different complexity based on the proximity, or visual importance of characters
to the camera.
Fuzzy controllers have a set of fuzzy rules, or a “rule base” that defines the inference of the controller.
There is no assurance that hand-calibrated rule bases are optimal, and indeed that calibration
based on fixed test environment will apply well to a dynamic environment. Special consideration
is given to genetic-fuzzy systems, which use a genetic algorithm to automatically calibrate a rule
base. Various architectures for genetic-fuzzy system are proposed and evaluated including dynamic
systems, which have the ability to learn “on the fly”, rather than in fixed experiment scenarios. A relationship
between genetic algorithm parameters and time-efficient fitness improvement is found.The
time requirements of training more complex “cascading” fuzzy systems are discussed. Distributed
and parallel training models are also considered.
A new, modular agent middleware is proposed, which is the underpinning software that perceives
the complex environment, feeds inputs into the fuzzy motion controllers, and effects output actions
for each character and vehicle. The middleware model is successfully used to drive a range of
vehicles and characters used in experiments.
The problem of evaluating motion controllers within a scientific framework is discussed. Several
candidate solutions are used, and a system for objectively evaluating mechanically simulated
vehicle motion is defined and evaluated. A complete tool-chain for designing complex simulations
and doing scientific experiments with them is is developed and discussed in detail, including simulation
software design methods, libraries, visualisation tools, and useful algorithms, a well-defined
mechanical simulation system, and practices for collecting data from simulations, and quantifying