Spaces:
Running
Running
File size: 2,785 Bytes
8634dd6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 | {
"schema_version": 1,
"title": "Reproduction: Corruption-Tolerant Asynchronous Q-Learning with Near-Optimal Rates",
"emoji": "🛡️",
"space_id": "snaykey/repro-async-q-learning",
"paper": {
"openreview_id": "L5ZJv73k7x"
},
"tags": [
"icml2026-repro",
"paper-L5ZJv73k7x"
],
"updated_at": "2026-07-30T16:28:22+00:00",
"root": {
"slug": "index",
"title": "Reproduction: Corruption-Tolerant Asynchronous Q-Learning with Near-Optimal Rates",
"file": "pages/index.md",
"children": [
{
"slug": "claim-1-theorem-2-rate",
"title": "Theorem 2 shows Robust Async-Q achieves a finite-time convergence rate of O(1/√T), matching classical asynchronous Q-learning, plus an additive O(√ε) term where ε∈[0,1/2) is the fraction of adversarially corrupted rewards (Section 3.1).",
"file": "pages/claim-1-theorem-2-rate/page.md",
"children": []
},
{
"slug": "claim-2-theorem-3-lower-bound",
"title": "Theorem 3 gives an information-theoretic lower bound of Ω(√ε), proving the additive corruption term in the upper bound is unavoidable (Section 4).",
"file": "pages/claim-2-theorem-3-lower-bound/page.md",
"children": []
},
{
"slug": "claim-3-theorem-4-reward-agnostic",
"title": "Theorem 4 extends the guarantee to a reward-agnostic setting where the algorithm needs no prior knowledge of the reward distribution's bounds (Section 5.1).",
"file": "pages/claim-3-theorem-4-reward-agnostic/page.md",
"children": []
},
{
"slug": "claim-4-theorem-6-markovian",
"title": "Theorem 6 extends the near-optimal finite-time rate to single-trajectory Markovian sampling with time-correlated data, with the bound inflated by the chain's mixing time (Section 6).",
"file": "pages/claim-4-theorem-6-markovian/page.md",
"children": []
},
{
"slug": "claim-5-algorithm-1",
"title": "Algorithm 1 (Robust Async-Q) combines a trimmed-mean reward estimator with an adaptive threshold G_t to filter corrupted rewards before the standard Q-learning update, requiring only knowledge of the corruption fraction ε and minimum state-action visitation probability λ_min (Algorithm 1, Assumption 1).",
"file": "pages/claim-5-algorithm-1/page.md",
"children": []
},
{
"slug": "executive-summary",
"title": "executive-summary",
"file": "pages/executive-summary/page.md",
"children": []
},
{
"slug": "conclusion",
"title": "conclusion",
"file": "pages/conclusion/page.md",
"children": []
}
]
}
} |