10 likes | 179 Views
Fault-Tolerant Computing with N-Version Genetic Programming Kosuke Imamura & James A. Foster Initiative for Bioinformatics and Evolutionary Studies (IBEST) Computer Science Dept. University of Idaho Moscow, ID 83844-1010 {kosuke,foster}@cs.uidaho.edu. Hypothesis
E N D
Fault-Tolerant Computing with N-Version Genetic ProgrammingKosuke Imamura & James A. FosterInitiative for Bioinformatics and Evolutionary Studies (IBEST)Computer Science Dept. University of IdahoMoscow, ID 83844-1010 {kosuke,foster}@cs.uidaho.edu Hypothesis GP can reduce the mean time between failures of a system by providing modular redundancy in an N-version fault tolerant system. Path to be predicted So many GPs for the same training data ! N-version GP (NVGP) We developed a redundant module software system with an isolated island model GP. A fundamental assumption of N version programming is that independent modules will have uncorrelated faults, so that a composite system will be more reliable by avoiding simultaneous faults in different modules. Experiment Our NVGP consists of 5 GPs and one master. The master averages each GP’s prediction at time ti to predict the next location of a moving object at time ti+1. After the prediction is output, the master receives the actual location from an external source as feedback. If a GPs prediction error is beyond a pre-determined threshold, the GP is retrained using the 5 most recent actual locations received by the master. NVGP Single GP Prediction Error Plot Reliability Increase Error Threshold 0.04 0.03 0.02 0% 50% 100% 150% 200% 250% 300% 350% • Conclusions & Future Research • NVGP reduces variance and expected MTBF in system output • Speciation strategies for evolvable systems (better fault masking) • Real-time self-epairing Evolvable Hardware (GP Pooling)