train

Train

{Train}=require('flatnet')

Simple training function: Evaluate, Fitness, Save?, End?, Cycle, repeat

• Train(net,data=[],targets=[],config=CONFIG)
• data array:
  • Elements in the array can be any variable, they will be Flattened
  • Each element should contain all the inputs for a single step evaluation
• targets array:
  • Elements in the array can be any variable
  • Elements are passed to the user supplied Fitness function
• config object:
  • name
    • The name used for file saving and loading
    • Default name = 'model'
  • keep_ratio
    • The % of models that will survive removal, see Network.Cycle
    • Default keep_ratio = 0.25
  • mut_count
    • The number of times the model is mutated, see Network.Mutate
    • Default mut_count = 3
  • mut_ratio
    • The pivot point for deleting or adding actions in mutation, see Network.Mutate
    • Default mut_ratio = 0.4
  • logging
    • Console.log information during training
    • Default logging = false
  • Fitness
    • The fitness function used after evaluation
    • Default Fitness = Network._STDFITNESS, see Network._STDFITNESS
  • Save
    • This function is called after Eval & Fitness have completed, and before End and Cycle
    • To initiate a map save, must return a valid index position of a model in net.models[]
    • See Model.SaveMap
    • Default save:

       Save:func=(net)=>{
       	if(func.last>=net.scores.high)return;
       	func.last=net.scores.high;
       	return net.scores.h;
       }
      

  • End
    • The function to determine end of training
    • Is called after Save and before Cycle
    • Return === true to end
    • Default End = (net)=>false (infinite)