Replacing eligibility traces

⚠️ Curated Limitations & Negative Results

• This source-scoped recurrence is not the entire eligibility-trace family, a TD or Sarsa controller, MCQ-L, Peng–Williams Q(lambda), or a true-online trace algorithm. The original accumulating recursion already has a separate maintained TD-family context. General eligibility-family organization and other trace variants remain outside this record’s scope; the source’s earlier historical citations do not establish independently verified first-use priority.
• The [0,1] trace bound requires the stated initialization and gamma*lambda domain. A bounded trace does not bound arbitrary rewards, TD errors, step sizes or parameter changes, and does not itself prove stable or convergent learning. The endpoint gamma*lambda=1 retains inactive traces rather than forgetting them; no continuing undiscounted-return existence result follows from it.
• The first-visit correspondences retain the absorbing-process, completed-trial and step-size conditions stated in the claim. The batch result is a convergence limit, not equality after an arbitrary finite number of sweeps, and alpha=0 is excluded. The paper’s MSE comparisons are conditional and non-strict, with equality cases; they are not a universal superiority claim. No complete appendix-proof or external convergence-theorem certification is asserted.
• The selected chosen-action feature replacement with clearing of other actions is one control realization, not a universal definition of replacing traces. The source explicitly leaves comparisons with state-action-only replacement open. Tabular and binary-feature results are not a general nonlinear-network replacement prescription, and stationary prediction results do not automatically apply while the control policy changes.
• Figure 9 initializes all traces in step 1, but its end-of-trial branch returns to step 2 without an explicit all-trace reset. The note that the time index continues across trials does not settle that omission. This description preserves the printed recurrence and selected ongoing control steps without silently adding a reset or claiming a defect in unseen author software.
• Appendix B states that position is clipped within [-1.2,0.5] while also describing termination when the next position exceeds 0.5. A literal test after that clipping cannot terminate by this condition. Its displayed minus-g force together with negative g also requires a sign/convention clarification. These ambiguities limit a literal reading of that experimental environment; they do not invalidate the trace recurrence or establish how the experiments were implemented.
• The inspected evidence is a 36-page author-hosted prepublication layout associated with the 1996 journal article; local PDF page locators are used instead of inferred journal pagination. OCR was needed for prose, with selected equations and pseudocode checked visually. No author implementation, benchmark or full original-proof reproduction was performed for this evidence assessment; the reported random-walk and CMAC experiments do not establish universal performance or an optimal trace setting.