Engineering: Incorporating MLP into Detector

• Mean squared error is not appropriate
  • does not represent desired result
• Token Error Rate is not differentiable
  • cannot use back propagation
• Solution - two stage process
  • Start with back propagation on mean squared error
  • Continue with gradient descent on TER
  • Simulated annealing to avoid local extrema

• Validation - reproduce results from NICO implementation
• Gradient descent - recursive step size
• Annealing - add random dither to weights and repeat
• Results - best point gives 11.7% TER