Reference · The Applied Higher-Order Functions catalogue

rainbow.algorithms.proofs.ahof.behavioral-cloning

The Applied Higher-Order Functions cataloguesrc/rainbow/algorithms/proofs/ahof/behavioral_cloning.clj1 definitions, 4 theorems

AHoF D64 Behavioral Cloning (imitation learning): train a policy by supervised learning to reproduce an expert's actions, minimizing an imitation loss. The clone is accepted on a demonstration when its loss is within a tolerance: within-tolerance? loss tolerance = decide (loss <= tolerance). A zero loss (perfect match) is within any tolerance; a loss 1/4 within tolerance 1/2 is accepted; a loss 3/4 exceeding the tolerance is rejected. Proved native and axiom-free (decision, perfect and acceptable instances, at-most certificate, rejection). Requires audit/decide-nnrat-at-most, nnrat/at-most, nnrat/decide-at-most-true-sound (all in the D63 checkpoint). Scope (measured by the kernel): this module has 4 worked instances (statements about specific values). None of its theorems is a guarantee about the algorithm for all inputs; `lein verify rainbow.algorithms.proofs.ahof.behavioral-cloning` prints this classification.

cloning/within-tolerance? :
  (x₁ x₂ : ℚ≥0)
  → data/Bool

cloning/perfect-match-accepted :
  → search/poll-improves 0 1/2 = data/true

cloning/close-clone-accepted :
  → search/poll-improves 1/4 1/2 = data/true

cloning/acceptance-certificate :
  → 1/4 ≤ 1/2

cloning/poor-clone-rejected :
  → search/poll-improves 3/4 1/2 = data/false