An adaptive satellite-scheduling algorithm recorded the highest mean observation benefit in all 14 simulated maritime moving-target scenarios, according to a preprint. Its reported advantage over conventional ant colony optimization ranged from 3.40% to 9.40%. The comparison involved five algorithms, all run under the same scheduling model and evaluation budget.
The approach, called IQACO, uses offline implicit Q-learning, or IQL, alongside ant colony optimization. It first collects transitions, or records of earlier search decisions, from ACO runs to train a policy. That policy is then used while schedules are being built to adjust the search parameters, allowing the settings to change during construction rather than remaining fixed from start to finish.
The objective combined normalized observation benefit, energy efficiency and workload balance. In other words, the algorithm was searching against several competing priorities within one objective rather than optimizing a single measure alone.
How the comparison worked
The test set contained 14 scenarios with 100 to 240 maritime moving targets and three to six satellites. Each target had two to four observation requirements, which were expanded into scheduling tasks. Because these were modeled scenarios, the findings concern how the algorithms performed in simulation.
IQL training used separately generated scenarios with the same parameter ranges but different target distributions and initial satellite conditions. The 14 scenarios used for testing were reserved for that purpose, so the training cases and reported test cases were not the same.
GA, PSO, WOA, conventional ACO and IQACO each ran 20 independent times per scenario. Every method used the same model, objective, feasibility checks and evaluation budget, with a common stopping criterion of 20,000 NFE. That setup kept the algorithm comparison on matched terms.
The gap widened in larger cases
Across the full 14-scenario set, IQACO's mean observation benefit was highest every time. In the repeated-run results, its mean benefit was 0.6564 with a standard deviation of 0.0167 in Scene 01, and 0.6019 with a standard deviation of 0.0120 in Scene 14. The paper reports IQACO as significantly better than every baseline at p < 0.05.
The difference was especially visible in the larger cases. In Scenes 10, 12 and 14, IQACO reached a higher best-so-far objective level than the competing algorithms, with greater separation as the number of targets and satellites increased. Best-so-far refers to the strongest result found up to a given point in the search, so the comparison also shows how the algorithms progressed.
The researchers also changed the relative weights assigned to the objective's three goals. On Scenes 01, 04, 07 and 11, they tested five weight configurations with 10 runs each. IQACO produced the highest or near-highest values under most configurations, with a stronger advantage when the objectives were balanced.
Training diagnostics showed IQL losses falling early and then fluctuating within bounded ranges without evident divergence. Mean Q- and V-value estimates became more stable later in training. The supplied analysis does not provide the numerical values behind those training curves.
What the simulations leave open
The boundaries of the evidence are important. The authors say the study relies on simulated target distributions, simplified cloud-availability modeling and an offline policy trained within a fixed scenario distribution. That leaves open whether the same advantage would hold with different target data, cloud conditions or operating constraints.
The separation between training and testing does not remove that uncertainty: the policy was trained within a defined scenario distribution, even though the test scenarios had different target distributions and satellite initial conditions. The reported gains should therefore be read as results from the tested model and parameter ranges, not as a general guarantee.
This is an arXiv preprint identified as version 1 of arXiv:2608.24471 [cs.AI], dated 25 August 2026. No funding statement is reported. The authors declare no known competing financial interests or personal relationships that could have influenced the work. Supporting data and code are available from the corresponding author upon reasonable request.
Paper data and sources
Original title: Implicit Q-learning-bootstrapped ant colony optimization for maritime moving-target observation scheduling with agile satellites
Authors: He Wang, Junyu Wu, Yeye Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text