Hierarchical Distributed Genetic Algorithms: A Fuzzy Logic Controller Design Application

• Authors: Jinwoo Kim;Bernard P. Zeigler
• Venue: IEEE Expert: Intelligent Systems and Their Applications
• Year: 1996

Unlike conventional genetic algorithms, HDGAs use a multiresolutional search scheme and change structure to achieve a goal. In this application, an HDGA investigates various configurations dynamically to find the optimal design of a fuzzy logic controller for a complex, unmanned, space-based plant.With the emergence of high-performance computing power, design engineers have applied AI techniques to a wide spectrum of real-world problems in intelligent and autonomous control.1 At the University of Arizona, for example, we have developed a fuzzy logic controller (FLC) to control the temperature of a space-based plant designed to produce oxygen from the Martian atmosphere.2 The controller must operate in two modes. In automatic mode, it increases temperature from a fixed initial point to a fixed set point at a fixed rate. Designing a conventional PID (proportional integral derivative) controller for these operational specifications is routine. In autonomous mode, however, the control system must cope with abnormal situations that can shut down plant operation. Recovery from such conditions involves reheating from arbitrary initial points at different rates and possibly to different set points. A single PID cannot handle all these situations efficiently, and manually designing an FLC to satisfy such requirements is virtually impossible. Instead, we used the AI technique of genetic algorithm optimization to design an FLC that satisfies these multiple operational specifications.Conventional von Neumann computers are not appropriate for executing AI programs because their architectures are designed for sequential and deterministic numeric computation.1 AI codes for real-world applications are highly complex and heterogeneous, and researchers have made many efforts to develop suitable AI computing architectures. As a conceptual framework, we have developed the Intelligent Machine Architecture (IMA), which integrates nondeterministic symbolic processing and computationally intensive numeric processing. To solve complicated problems demanding large amounts of data, knowledge, and processing, the IMA distributes processes in layers according to their computational characteristics. This multilayer interconnection scheme provides a high degree of parallel processing.Within the IMA framework, we developed the hierarchical distributed genetic algorithm (HDGA). Genetic algorithms (GAs), which use the principles of biological evolution, often outperform classical optimization methods in solving complex design problems. Because they simultaneously evaluate many points in the parameter space, GAs are more likely to converge toward a global solution.4 However, real application problems often require a model with many parameters. To achieve optimal results, the algorithm must tune each parameter to a high degree of precision. These parameters not only increase the problem's complexity, they also exert varying degrees of influence on system performance. In existing approaches, a chromosome, or string of search parameters, does not contain information about the parameters' performance impact. To address this problem, the HDGA supports a multiresolutional search paradigm and employs a variable structure-that is, it changes its internal structure to achieve a goal.