the l-co-r co-evolutionary algorithm: a comparative analysis in medium-term time-series forecasting...
TRANSCRIPT
Presentation Template Design-7
The L-Co-R co-evolutionary algorithm: a comparative analysis
in
medium-term time-series forecasting problems
Parras-Gutirrez, Rivas and Merelo
U. Jan & Granada (Spain)
http://geneura.wordpress.com
It's difficult to make predictions, especially about the future
Yogi Berra
Cropped from Image by Pensiero at http://www.flickr.com/photos/pensiero/2878055175/
Smells like a bubble
Using coevolution to predict bubble-bursting
Image by Light Knight at http://www.flickr.com/photos/lightknight/3176554248
Radial Basis Function neural nets and time lags
Coevolving!
Image by JWPotowerks at http://www.flickr.com/photos/john_whitworth_photography/3017645549 (spokes) and Eduardo Zrate at http://www.flickr.com/photos/eduardozarate/3513912756/
What are RBFNNs?
What do we mean by time lags?
Horizon is what lies between the predicted value and the first previous datum used to predict it. A consistent value of horizon was used throughout the experiments.
Trend pre-processingTrend post-processingInitializate lagsInitializate RBFNNEvaluate lagsEvolve Lags: CHCEvaluate RBFNNEvaluate LagsEvolve RBFNN: EAEvaluate RBFNNRBFNNsLagsMain loop
Lags' loop
RBFNs' loop
Final forecasting
CHC combines conservative selection strategy with disruptive recombination HUX: http://neo.lcc.uma.es/mallba/easy-mallba/html/algorithms.html#chcCHC is called also Adaptive Search AlgorithmEvRBF is an already published evolutionary algorithmTrend is removed and then added to avoid it to dominate the prediction.
Let's fight
Data sets taken from Spanish National Statistics Institute+ Time Series book by D. Pea + NN3 competitionCheck them out at https://sites.google.com/site/presetemp/datos
Airline passengers, mortgages, prices...
Comparison with other five methods:Exponential Smoothing Method.
Croston
Theta
Random Walk
ARIMA
75% for training - 25% for testing30 executions with average published.
L-Co-R predicting airline passengers
How do we measure success?
Several measures used:Mean absolute percentage error : MAPE.
Mean absolute scaled error: MASE.
Median absolute percentage error: MdAPE.
MASE is probably the most reliableLess sensitive to outliers.
Less variable on small samples.
More easily interpreted.
Imagen by tudedude at http://www.flickr.com/photos/tudedude/3516187441
Who's the best?
Differences are significant anyways
That's all
Any questions?
Check us out at @geneura@canubeproject@anyselfproject@sipesca
ANYSELF
AnyselfProject