/toycsw

toy models for "CoffeeShop World" experiment: effect of training training curriculum on schema learning in humans and neural networks

Primary LanguageJupyter Notebook

main: handheld splitting + schema rnn

  • splitting controlled by RNN
    • 04/19/21: found test B>I without much optimizing.

working notes:

  • analysis of best fit handheld splitting + schema RNN
    • plot splitting
    • best fit:
      • I more splits
      • I never enters pe<thresh
      • B more pr_stay
  • can I specify nonprobabilistic splitting procedure?
    • something such that, when active schema is below PE thresh, high weight given to current schema.

notes

  • placing pe thresh condition as default (first) schema selection condition also yields good B>I.
    • good news because now pr_stay only applies when pe_thresh not met
  • best fit: {'sticky_decay': 0.02, 'pe_thresh': 0.9, 'init_lr': 0.45, 'lr_decay': 0.25, 'stsize': 6.0}
  • stsize 10 yields poor fit (I learns too fast)

delta+rnn branch

  • splitting controled by delta learner, performance evaluated on rnn
    • good B>I without much optimizing