Venue 2 of 6 · Automated planning · Archival conference

ICAPS 2027: Main Track

The International Conference on Automated Planning and Scheduling, Columbia, South Carolina, 27 June to 2 July 2027. This is the flagship planning venue, and the natural home for the utility function as a planning contribution.

Paper due 14 Dec 2026 AoE8 pages + refs · AAAI styleDouble-blindFacts checked 11 Oct 2026
Needs Phase 2 test resultsTight: 9 weeks

Strong topical fit. The call lists planning under uncertainty with “incomplete models or information” and “belief states” as topics, and that is the open-world h₀ term exactly.

But ICAPS is a results venue. Without a preregistered P-vs-B3 comparison at ≥1,000 events per scenario, reviewers will reject it. Go only if Phase 2 test-seed results exist by about 1 December 2026. Otherwise, target ICAPS 2028 with the full study.

Venue facts

ThemeAutomated planning and scheduling. Relevant listed topics: planning under uncertainty (MDPs, POMDPs, planning with sensing, incomplete models, belief states), real-time and online planning, novel application domains. A special track, “Planning Under a Different Name,” targets neighbouring communities.
FormatLong paper: 8 pages + references. Short paper: 4 pages + references. AAAI format mandatory. Content past page 9 (long) or 5 (short), other than references or the ethics statement, means desk rejection.
ReviewDouble-blind. No names or affiliations, own work cited in the third person. arXiv posting discouraged from two weeks before the deadline until notification.
Deadlines (AoE)Abstract 7 Dec 2026 · paper 14 Dec 2026 · notification 26 Feb 2027.
SourcesICAPS 2027 CfP · ICAPS 2027 home

Tailored abstract

Open-World Information-Gain Planning with Decaying Measurement Value

We study sequential measurement planning where the goal is to identify which of K generative models explains an evolving process, where the informativeness of each candidate measurement decays at a model-dependent rate, and where the true model may lie outside the set. Standard expected-information-gain planners assume a closed hypothesis set and are indifferent to when information becomes unrecoverable, so they can commit confidently to a wrong model and spend budget on measurements that could have waited. We propose a planner whose utility adds two terms to closed-world class information gain: the mutual information between a measurement and a model-adequacy indicator, computed against a Gaussian-process auxiliary hypothesis, and an opportunity-decay term that values a measurement by the information lost if it is deferred to the next feasible slot. When inadequacy becomes probable, the planner switches from discrimination to evidence preservation. We evaluate on fading astronomical transients observed by a simulated low-Earth-orbit constellation with SGP4 visibility, against fixed-cadence, nearest-capable, classifier-entropy and closed-world EIG baselines under identical budgets and preregistered test seeds. Result sentence pending: time-to-identification and AUROC of P(h₀) from Phase 2 test seeds.

Paper plan (8 pages, AAAI two-column)

  1. Introduction0.9 pFrame it as a planning problem first (open-world sequential model discrimination) and the astronomy second, as the motivating domain. Three contributions: the utility, the mode switch, the benchmark with withheld classes.
  2. Related work0.7 pPlanning-centric: active hypothesis testing, POMDP information gathering, robust/misspecified Bayesian experimental design, POISE, SHM-POMDP. Astronomy follow-up in one paragraph.
  3. Problem formulation1.0 pEqs. 1–2 (hypotheses, GP h₀) and the feasible-action model. Define the decision problem formally as a belief-MDP with a deadline-dependent observation model.
  4. The open-world decaying-value planner1.4 pEqs. 3–4. State honestly that D is currently a Fisher-style SNR² proxy, not eq. 4's literal per-class MI (see research/LOG.md). Either implement the exact term or present the proxy as the method with a justification. Mode switch, and the exact quadrature EIG estimator with its known-answer tests.
  5. Experimental setup1.0 pClasses, constellation, B0–B3, metrics, scenarios (Nominal, Crowding, Model error). Seeds and the preregistration. The β, γ, τ calibration protocol on validation seeds.
  6. Results1.8 pGenerated only from research/results/phase2/. Headline: H1 (time to 95%) and H3 (AUROC of P(h₀) vs. entropy threshold), with Wilcoxon, Holm and bootstrap CIs. Ablation: γ=0 split by fast vs. slow classes (H2). A refuted hypothesis is reported as refuted.
  7. Discussion and limitations0.7 pAnalytic light-curve approximations, uniform cost, false alarms under model error. Decentralisation (H4) is future work. Don't overreach into it in this paper.
  8. Conclusion0.3 pOne paragraph.

Before you submit

← AAMAS Blue Sky · All venues · Next: IWPSS →