Computer-aided pipeline operation using genetic algorithms and rule learning. PART II: Rule learning control of a pipeline under normal and abnormal conditions

In this two-paper series, techniques connected with artificial intelligence and genetics are applied to the problem of gas pipeline control. In the first paper, genetic algorithms were applied to two pipeline optimization problems. In this, the second paper, genetic algorithms are used as a basic learning mechanism in a larger rule learning system called a learning classifier system. The learning classifier system is developed and applied to the control of a gas pipeline under normal summer and winter operations as well as abnormal operations during leak events.A learning classifier system is a software system that learns rules called classifiers to guide its behavior in arbitrary environments. Environmental information comes in through sensors and decodes to a finite length message. Messages on a message list match and fire rules called classifiers with explicit pattern recognition capability. Classifiers, once matched, send messages to the message list. These messages may in turn match other classifiers, or they may fire action triggers called effectors. Using environmental reward, the system selects good rules through a reward allocation system based on a competitive service economy. Competition encourages the survival of good rules, those that set up environmental reward. Furthermore, the system tries and learns new rules using a genetic algorithm similar to the one presented in the previous paper.Together, the learning classifier system with its complete rule and message system and powerful learning heuristic is capable of learning how to operate a pipeline under normal and abnormal conditions alike. Computational experiments are presented that demonstrate the systems learning ability under summer and winter conditions starting from a random state of mind. These results compare favorably with a random walk through the decision space. Additionally, the learning classifier system is trained to detect leaks. Repeated exposure to simulated leak events results in the development of rules that permit a high percentage of detected leaks and a low percentage of false alarms.