Online hybrid temporal regression planning

⚠️ Curated Limitations & Negative Results

• The papers attribute optimality to the search for the current job under its objective and existing-plan context, relying on heuristic admissibility and suitable feasible-state handling. Their greedy outer loop does not establish a globally optimal schedule over future arrivals, a competitive ratio, or optimal revision of earlier action choices.
• The heuristic prose labels a previously unknown literal when produced, but does not give an explicit agenda or label-improvement rule ensuring minimum relaxed achievement times. Admissibility cannot be inferred for an arbitrary first-producer implementation; a fast and a slow producer can require different label handling. This is an exposition/premise limit, not an observed implementation failure.
• Earliest-time clamping is justified only with the source’s appropriate complete temporal-window maintenance and its domain assumptions. It is not a generic guarantee for arbitrary CSPs, approximate or stale windows, or an unspecified propagator. The comparison does not demonstrate that a correctly maintained simple temporal network becomes inconsistent merely by earliest-time clamping.
• The manually chosen imminence interval assumes sufficient time to plan the next job; the papers discuss interrupting and restarting when needed. Outer-loop temporal updates are restricted to between-job searches, while controller changes can abort a current search. These descriptions do not supply a general hard real-time guarantee or fully specified timeout, unreachable-job and empty-frontier behavior.
• The state-uniqueness argument appeals to irreversible logical or resource-order choices reflected in a plan. Local branching distinctions alone do not establish global semantic non-reconvergence, cycle freedom, or a general no-duplicate theorem for arbitrary action domains. No concrete duplicate execution of the authors’ implementation is asserted.
• The 2005 simplified rule tests only whether effects match a fully specified state. Full specification plus corresponding effect preconditions does not alone ensure compatibility of preconditions on unaffected atoms. Such a precondition can contradict the target state even when the simplified effect test passes. The earlier general applicability rule retains the needed check; no implementation defect or author erratum is established.
• The 2005 simplified update flips added/deleted atoms. That update needs opposite-polarity preconditions for the changed atoms and suitable treatment of other preconditions; the phrase corresponding preconditions does not itself specify those polarities. The general regression update is retained separately, and the simplified branch is not silently strengthened or treated as universally equivalent.
• The 2005 paper explicitly supports unknown literals and describes storing true and unknown sets, but later calls the state fully specified and writes a two-set heuristic shorthand. The exact treatment of unknowns in that shorthand remains a source/interface question; this draft does not erase unknowns or claim an implementation-level equivalence of all these descriptions.
• The shared procedure is described for the supplied symbolic manufacturing model: jobs are independent except for shared resources and explicit batch constraints, actions for one job are sequential and abut, and controller delay constrains starting times. These restrictions do not establish applicability to arbitrary interacting, stochastic, partially observable or general durative-planning domains.
• Representational flexibility is not the same as implemented support: the 2004 paper allows duration ranges but says its implementation uses fixed durations, and it describes additional resource-state compatibility restrictions. These details and the 2005 formal extensions remain version-local; this common record does not equate every modeling option or implementation across the two papers.
• The described failure handling assumes reliable transport along planned routes and destination diversion that can discard faulty material; affected jobs can then be replanned from scratch. This does not establish recovery for arbitrary mid-route failures, uncertain state estimation or all in-flight material. Controller-domain changes are handled by aborting the current search under the stated between-search model.
• The 2004 experiments distinguish incremental directed path consistency from an arc-consistency variant maintaining origin-relative time windows without adding arcs. The latter does not provide arbitrary pairwise time relations as easily. The comparison does not prove interchangeability of arbitrary AC-3 and IDPC implementations or transfer the reported timing advantage to every domain.
• Physical-prototype operation, simulations on selected plant configurations and timing/quality comparisons are separate evidence roles. They do not establish universal throughput, hard deadlines or general optimality. This draft makes no numerical performance claim and has not executed an implementation or reproduced the experiments.
• One descriptive identity covers the common outer online coordination and inner temporal/resource regression procedure in these two inspected works, not separately originated A*, STRIPS, STN or propagator components. It does not equate all later printer-planning publications or use the filename token pns as an alias. The 2005 comparison is completed, while remaining cited-original, later-version and implementation questions stay separately tracked.