machine-earning / augmented-population-based-training Goto Github PK

how to implement early stop or overfitting prevention

• random weight initialization
• variable number of layers or variable topology
• variable unit types (much later)

• make sure PBT are working properly (use debugger)
• find best tune
• set the best as the most accurate, not the highest performance
• find the right evaluation function that prioritizes the accuracy but restrains the size just enough to produce high accuracy, low size networks

• Tennis
• Iris
• Identity

• remove layer of single unit
• remap network
• update labels

• implement the functions
• architecture search - number of layers and number of hidden units per layer
• encode architecture