Dynamic System Control Using Rule Learning and Genetic Algorithms

In t h i s paper, recent research resu l t s are presented which demonstrate the ef fect iveness of a ru le learn ing system in two dynamic system con t ro l tasks. This system, ca l led a learn ing c l a s s i f i e r system (LCS), learns ru les to con t ro l a simple l n e r t l a l object and a simulated na tu ra l gas p i p e l i n e . S ta r t i ng from a randomly generated s ta te of mind, the learn ing c l a s s i f i e r system learns s t r i n g r u l e s ca l l ed c l a s s i f i e r s which match s t r i ngs ca l led messages. Messages are sent by environmental sensors or by prev ious ly ac t i va ted c l a s s i f i e r s . Each c l a s s i f i e r ' s e f fect iveness i s evaluated by an i n t e r n a l serv ice economy complete w i th b idd ing and auc t ion . Furthermore, new ru les are created by an innovat ive search mechanism ca l led a genetic a lgor i thm. Genetic algor i thms are search algor i thms based on the mechanics of natura l genet ics . Results from computational experiments in both tasks are presented. In the i n e r t i a l object task , the LCS learns an e f f e c t i v e set of ru les to center the object repeatedly. In the p ipe l i ne task , the LCS learns to con t ro l the p ipe l i ne under normal summer and w in te r cond i t ions . It a lso learns to alarm co r rec t l y fo r the presence or absence of a leak. These resu l t s demonstrate the ef fect iveness of the learn ing c l a s s i f i e r system approach and suggest f u r t h e r refinements which are cu r ren t l y under I n v e s t i g a t i o n .