Theorem 2.3: k-decision lists are PAC learnable by the greedy decision-list algorithm, which never fails and whose outputs are consistent
ProvedComputationalLearning.decision_lists_pac_learnableTheorem 2.3. For any fixed , the representation class of -decision lists is efficiently PAC learnable.
Formally, for the greedy decision-list algorithm of p. 44 (repeatedly append a useful condition with its label and remove the examples it satisfies, until none is left), with conditions the conjunctions of literals ( of them): (i) on every sample labeled by a -decision list, every partial run with examples remaining can be extended by a useful condition (the algorithm always succeeds); (ii) every complete output is consistent with its sample; (iii) for every target -decision list, distribution on , , and with , the probability that some output of the algorithm has error greater than is at most (an output uses each condition at most once, so the hypothesis class has at most members and Theorem 2.2 applies); (iv) -decision lists are PAC learnable using -decision lists. The book proves and reduces general to it by expansion; here the conditions are the -conjunctions directly, so the hypothesis is a -decision list.
n ≥ 1. The conditions are conjunctions of exactly literals, fewer by repetition, so for there is no empty (always true) condition. When no condition of the target is satisfied by a remaining example, every remaining example carries the target's default bit, and then any condition satisfied by a remaining example is useful; such a condition exists as soon as there is a variable. With and there are no conditions at all, and the greedy algorithm would fail on every nonempty sample, so the statement takes the book's (variables ). The book's own proof, "the first condition in the target decision list such that is non-empty is a useful literal", has the same gap when every remaining example falls through to the default; the argument above closes it.
import Definitions.Def_ComputationalLearning_Occam open MeasureTheory
namespace ComputationalLearning
/-- **Theorem 2.3** (p. 43). For any fixed `k ≥ 1`, the representation class of `k`-decision lists
over `n ≥ 1` variables is (efficiently) PAC learnable, by the greedy decision-list algorithm: on a sample labeled by a
`k`-decision list the algorithm always finds a useful condition while examples remain (so it
never fails), every output is consistent with the sample, and with
`m ≥ (1/ε)(ln N + ln(1/δ))` examples, `N = 2(2M + 1)^M` for `M = (2n)^k` the number of
conditions (an output uses each condition at most once, so it is one of at most `N` lists), the
probability that some output has error greater than `ε` is at most `δ`; hence `k`-decision
lists are PAC learnable using `k`-decision lists. -/
theorem decision_lists_pac_learnable {n k : ℕ} (hn : 0 < n) (L₀ : DecisionList n k)
(D : Measure (Cube (Fin n))) [IsProbabilityMeasure D] {ε δ : ℝ} (hε : 0 < ε) (hε1 : ε ≤ 1)
(hδ : 0 < δ) (hδ1 : δ < 1) (m : ℕ)
(hm : 1 / ε * (Real.log (2 * (2 * (Fintype.card (Condition n k) : ℝ) + 1) ^
Fintype.card (Condition n k)) + Real.log (1 / δ)) ≤ m) :
(∀ S : Fin m → Cube (Fin n) × Bool, IsLabeledBy (evalDL L₀) S →
∀ (steps : List (Condition n k × Bool)) (R : Finset (Fin m)),
IsGreedyPrefix S steps R → R.Nonempty → ∃ (c : Condition n k) (b : Bool), IsUseful S R c b) ∧
(∀ (S : Fin m → Cube (Fin n) × Bool) (L : DecisionList n k),
IsGreedyOutput S L → IsConsistent (evalDL L) S) ∧
sampleLaw D (evalDL L₀) m
{S | ∃ L : DecisionList n k, IsGreedyOutput S L ∧ ε < errorOf D (evalDL L₀) (evalDL L)} ≤
ENNReal.ofReal δ ∧
PACLearnable (decisionListClass n k) (decisionListClass n k) := by sorry
end ComputationalLearning
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.