Under its queueing assumptions, a mathematical study constructs repeatable policies with lower average expected cost than fixed service when the cost of repair is below the critical saving from a single control period. The analysis denotes that positive difference as Δc, the critical gap.
The setting is an M/M/1 queue: arrivals are modeled as Poisson events, service times as exponential, and a queue-length sensor guides service-rate control. While the sensor is active, the policy chooses one of two rates, μ1 or μ2. After a breakdown, the system uses baseline rate μ, while repair is either immediate or scheduled for a future time.
How the modeled system works
The benchmark is a fixed-service policy that schedules no future repair. Costs combine service operation, queue holding and a per-repair charge; the model sets the cost of μ1 to zero and the cost of μ2 above it. Holding costs are assumed to be nonnegative, nondecreasing and convex.
The authors analyze the no-repair version as a discrete-time Markov decision process, then use value iteration and an average-cost equation. A Markov decision process is a way to represent choices as a system moves between states; here, the no-repair policy has a queue-length threshold form, and the candidate repair policies use threshold rules as well.
To extend a one-off control period into a repeating policy, the analysis combines renewal-reward and Tauberian theorems to link finite average expected cost with values from a calculation in which discounting of future costs fades.
The guarantee is mathematical
For the construction using μ1, the single-repair result gives a cost advantage of at least one-third of Δc as discounting of future costs fades, for every starting queue state. Its repeating version has an analytically guaranteed average cost no greater than the baseline minus Δc/[4(T + 1/β)], where T is the waiting time and β is the sensor-failure rate. The authors say this analytical lower bound may be conservative.
For the construction using μ2, the same one-third-gap guarantee applies only to starting queues no larger than a truncation level m. Larger queues are handled with a geometric tail bound. The corresponding repeating policy has average cost no greater than the baseline minus Δc/[12(1/β + Tm)], where Tm is the relevant waiting time.
Simulations suggest a pattern, with uncertainty
The computational check uses renewal-reward simulations. It estimates average cost by dividing mean cost accumulated over renewal cycles by mean cycle length, and reports approximate 99% confidence intervals.
The computational section reports three parameter cases with different modeled arrival rates, sensor-failure rates and service rates.
Across the μ1 runs, estimated savings generally increased as Δc grew, while the waiting time T1 decreased. The formal lower bounds were much smaller than the observed simulated savings, but the approximate 99% intervals could include the baseline cost, limiting how firmly the pattern can be read.
For μ2, the pattern depended on the case. Cases 1 and 2 showed a seemingly steeper increase in average cost reduction as Δc grew, while Case 3 did not show cost reduction as Δc increased; the authors attribute that result to high uncertainty. The truncation level m also decreased as Δc increased.
A model result with clear limits
These are model-specific results, not evidence from an operating service system. The study compares restricted threshold-policy constructions with a fixed baseline, and it does not establish global optimality across all possible repetitive repair policies.
The paper is an arXiv preprint, version 1 dated 20 August 2026, and the result is presented as a model-and-simulation finding under its stated assumptions.
Paper data and sources
Original title: To Control or not to Control
Authors: Odysseas Kanavetas, Camiel M. P. Koopmans, Floske M. Spieksma
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text