Synopsis

1
Synopsis

• Two data sets RO2A2281 (62 samples) RO2A2279
  (1750 samples)
• For RO2A2279, required modeling accuracy (average
  error lt 0.16, maximum error lt 0.8) can be
  achieved with 13 probes, but not with 9 probes
  (redundant probes removed)
• Applying RO2A2279 model to RO2A2281 data set
  fails to meet the accuracy requirement
• Characteristic of RO2A2281 data set is very
  different from RO2A2279
• Data too sparse
• High level of noise?

2
RO2A2279 Data Set With 13 Probes

• A network with 4 hidden neurons is enough to
  achieve accuracy requirement
• Training max error 0.78 average error
  0.12
• Testing max error 0.63 average error 0.12

3
RO2A2279 Data Set With 9 Probes

• Unable to achieve accuracy requirement
• Readings from redundant probes are not the same
• Data most likely contains readings from failed
  probes

4
RO2A2279 Model On RO2A2281 Data

• According to AIC measure, RO2A2279 model should
  have 8 hidden neurons
• Applying 8-neuron RO2A2279 model to RO2A2281 data
  set average error 1.45, maximum error 6.05

5
Analysis of RO2A2281 Data

• RO2A2279 model not applicable to RO2A2281 data
  set
• Characteristic RO2A2281 data set is significantly
  different from that of RO2A2279

• AIC (Akaiki Information Criterion) measures the
  goodness of a model (smaller AIC indicates a
  better model)
• Model with 1 hidden neuron has the smallest AIC
  (very unusual)
• Data too sparse
• Contains high level of noise?