Time Series Prediction with Mixture of ExpertsIntroductionImplementationTesting MethodologyResultsAR ModelExperts 1 ResultExpert 2 ResultExpert 3 ResultMixture of Experts ResultTable of ErrorsConclusionTime Series Prediction with Mixture of ExpertsA ECE539 Project By: Jiong FanIntroductionTime Series Prediction can be defined asy(t+1) = f(y(0), y(1), …, y(t-L))Times Series Prediction has a lot of applicationsThere exists a lot of method to perform this perdicitonImplementationThe implementation consists 3 experts and gating network Each of the expert is a MLP implementationExperts are chosen from a pool of predefined configurationsThe Gating Network is another MLP implementationTesting MethodologyTest file is from the class web page“Time-Series Prediction Competition”Data file is then divided into 2 partitionsThe last 270 is the testing dataThe rest is training and tuning dataAR model is used as base modelResults0 50 100 150 200 250 300-0.500.51Final predition vs. originalOriginal test dataMLP gatenetwork prediction0 50 100 150 200 250 300-0.500.51Expert(1) predition vs. originalOriginal test dataMLP expert(1)0 50 100 150 200 250 300-0.500.51Expert(2) predition vs. originalOriginal test dataMLP expert(2)0 50 100 150 200 250 300-0.500.51Expert(3) predition vs. originalOriginal test dataMLP expert(3)AR Model 1700 1750 1800 1850 1900 1950 2000-0.4-0.3-0.2-0.100.10.20.30.40.50.6y(t)yhat(t)Experts 1 Result0 50 100 150 200 250-0.4-0.3-0.2-0.100.10.20.30.40.50.6Expert(1) predition vs. originalOriginal test dataMLP expert(1)Expert 2 Result0 50 100 150 200 250-0.4-0.3-0.2-0.100.10.20.30.40.50.6Expert(2) predition vs. originalOriginal test dataMLP expert(2)Expert 3 Result0 50 100 150 200 250-0.4-0.3-0.2-0.100.10.20.30.40.50.6Expert(3) predition vs. originalOriginal test dataMLP expert(3)Mixture of Experts Result0 50 100 150 200 250-0.4-0.3-0.2-0.100.10.20.30.40.50.6Final predition vs. originalOriginal test dataMLP gatenetwork predictionTable of ErrorsAR Expert 1 Expert 2 Expert 3 Mixture of Experts0.016115 0.0014062 0.0059747 0.0026146 0.00070705ConclusionThe Mixture of Expert system predicts with more accuracy than any of the model
View Full Document