general workflow

• select ProtoRun from database
• store Run with chosen ProtoChromosome
• store generation 0
• randomly initialize and train generation 0
• train next generation
  • select, combine, and mutate
  • enqueue train requests containing:
    • generation ID
    • chromosome
  • collect train notifications
  • evaluate generation
    • evaluate Mixtures
    • attach MixtureEvals
  • attach GenEval

request workers take a Chromosome as input

• (an expert is the training output in addition to aux training input (e.g., initial weights) and metadata like final cost, cost over time during training, etc.)
• train expert
  • preprocess data
  • train neural net
• store Expert
• broadcast training notification

TODO

• web UI
• backtesting
• hotswap dlls isn't working properly. xcopy gives "Sharing violation"
• why is the GA overfitting the blind set? something about the mixtures must be degenerate
  • could have been the fact that I was accidentally using 30% testing set, 70% validation
• things to try
  • feed training set into experts to build up state before evaluating validation set?
  • experts vote either -1 or 1, or their vote is weighed by their certainty (magnitude of output)
• pick better initial weights: http://www.heatonresearch.com/encog/articles/nguyen-widrow-neural-network-weight.html
• catch NaN bug