Claims

The tool's Q-learning matches a value-iteration ground truth in 98.5% of non-terminal states on average

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims

Reading the evidence codes

q is the design tier, i the impact, n the sample, and the kind with its r the rigour judged within that kind: what each code means.

This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.

q2 i? Across 10 seeds, the greedy policy learned by the tool's Q-learning rule lies in the optimal action set for 98.5% of non-terminal states on average, with maximum absolute value error averaging 0.34. → Jueng 2026

Evidence

Jueng 2026

Jueng, Ekkachai. (2026). Q-Learning Lab: Teaching Reinforcement Learning Through Learner-Generated Trace Analysis. https://arxiv.org/abs/2607.10802 doi:10.48550/arxiv.2607.10802

q2 · i? · design · r2

Numerical validation on a Python reference implementation mirroring the JavaScript environment constant-for-constant: tabular Q-learning run for 50,000 episodes with a Robbins–Monro learning-rate schedule against value iteration converged to tolerance 10−12, scoring agreement by Eq. 2. The greedy policy matches an optimal action 'in98.5%of non-terminal states on average'.

"Across 10 seeds, the greedy policy matches an optimal action (Eq. 2) in98.5%of non-terminal states on average (minimum95%, i.e. at most one state per run)"

Discussion

Page metadata
Field Value
type claim
id q-learning-matches-value-iteration-ground-truth
title The tool's Q-learning matches a value-iteration ground truth in 98.5% of non-terminal states on average
description The tool's Q-learning matches a value-iteration ground truth in 98.5% of non-terminal states on average
status review
evidence_strength moderate
newest_evidence 2026
recheck_by 2029
generated by process:wiki-ingest on 2026-10-11
trust unverified — no one has confirmed this page's content
sources 1 — jueng-2026

Frontmatter is the page's machine-readable record — see the Schema & Guide.