An architecture for CMAC neural network adaptation of a genetic algorithm in a load balancing application

Limitations of the Genetic Central Task Assigner (GCTA) load balancing scheme are addressed. The GCTA scheme uses a genetic algorithm to allocate tasks among a collection of distributed processors. The effectiveness and adaptability of the GCTA is limited by the accuracy of the fitness function that is used to evaluate candidate task allocations. A novel architecture is introduced that augments the GCTA fitness function with a Cerebellar Model Articulation Controller (CMAC) neural network. The CMAC neural network is trained to predict errors in the GCTA fitness function based on experience. ne novel architecture uses the predicted error to improve selection of candidate task allocations. Preliminary results show an improvement in the performance of our architecture over the GCTA scheme as measured by a decrease in the average response time for tasks in the system.