Genetic learning of fuzzy control rules in hierarchical and multi-layer fuzzy logic systems with application to mobile robots

This thesis examines the genetic learning of fuzzy control rules. It applies concepts of evolutionary learning found in biology through the application of genetic algorithms and evolutionary algorithms to the complex task of rule determination in fuzzy control systems. A set of fuzzy rules is encoded as a chromosome of a genetic algorithm and the rules are learnt with no other knowledge than the fitness of each chromosome. This fitness is obtained by applying the fuzzy control rules to a specific problem application. The area of control of a mobile robot is chosen for application. The mobile robot is considered as point mass in the plane with kinematic equations of motion. Problems considered are control to a fixed target, control to a fixed target in the presence of obstacles, control of two mobile robots to fixed targets whilst avoiding collision, control of two mobile robots to two different targets whilst avoiding collision, control to capture a moving target.

These problems become progressively more complex in that the number of rules to be learnt increases at an exponential rate relative to the number of variables used. This problem is overcome by using multi-layer and hierarchical fuzzy logic system structures to reduce the number of rules to be learnt.

The multi-layer and hierarchical structure define a mean of interconnecting rule bases and improves the speed of learning of the control law as fuzzy rule bases for the fuzzy logic control system. These structures for concept learning using the integrated genetic algorithms and fuzzy logic makes easier the development of fuzzy logic control systems. It encourages the development of fuzzy logic controllers where the large number of system's parameters inhibits the construction of such controllers.

The control laws are learnt in hierarchical or multi-layer manner where the most important parameters of the system is used to construct the first layer rule base and the next most important parameters are used to develop the next layer rule base etc. The self-learning capability of the proposed system makes it able to learn the mapping between each layer of fuzzy rules in hierarchy. This creates the opportunity to apply this method to systems that need to be changed or modified as time passes.

Simulation experiments indicate that this approach can produce a control system that possesses a complex navigational capability through learning and adaptation to the environment workspace. These capabilities are a necessity for intelligent control systems.

The main achievement of this approach is that this architecture provides us with a general method to obtain knowledge in the form of fuzzy rules for the system to be controlled. The proposed architecture is a self-learning method, and it can quickly and effectively acquire the control knowledge and needs only the information about the performance of the system. Two methods for construction of fuzzy logic systems, namely multi-layer and hierarchical fuzzy logic systems are used with the proposed genetic fuzzy rule generator architecture to design and develop complex systems. The power of the proposed architecture combined with the multi-layer and hierarchical fuzzy logic structures are applied for solving problems in multi-robot systems.