Genetic Algorithms for Adaptive Real-Time Control in Space Systems

The U.S. Space Station planned for the 90's will comprise a large number of interacting control systems, a situation which will be too complex for humans to deal with without the aid of knowledge based systems. However, current knowledge based systems have two undesirable aspects. Brittleness: problem-solving performance degrades precipitously when the system is operating outside its intended domain, and adaptation: the ability of the knowledge base to deal with a changing environment. Learning systems try to alleviate these problems, by organising their own knowledge base. These systems can be divided into two categories: the more traditional rule-based systems on one hand,: and neural nets and genetic algorithms on the other. The latter two systems do not operate in the domain, but are "sandwiched' between a preand post processing phase. Systems based on genetic algorithms are characterized by on-thefly improvement through a continuous contention of the available rules, while maintaining performance levels as new rules are "tried out." The most salient feature is the employment of techniques from microbiology to conquer the combinatorial explosion involved in induction of the new rules.