KNOWLEDGE ACQUISITION AND ADAPTATION - A GENETIC APPROACH

A genetic algorithm-based (GA-based) system (GAPOLE) is used to evolve a self-learning, self-optimising control strategy for a typical inherently unstable, dynamic system - a simulated pole-cart system. The dynamics of the system are unknown to GAPOLE. The only information for evaluating performance is a signal indicating that the pole-cart system is out of control. This presents a genuinely difficult credit assignment problem We present some evidence which shows that GAPOLE compares well with the best alternative methods, but it is noteworthy that it is most robust. It is argued that maintaining a population of partial solutions offers some advantages: GA-based algorithms can deliver more than one good solution at a time; they are able to adapt better in complex changing conditions. Results characterising the performance of the method as population size is varied are also presented, The results show that GAPOLE performed best with a population size of 300. Therefore it is suggested that this parameter like other GA parameters, may have to be tailored to a particular application. This appears to contradict an earlier claim that a population size in the range of 60-110 is optimal for genetic algorithm-based applications (Grefenstette 1986) across domains.