A deep reinforcement learning framework for adaptive control charts
Sabahno, Hamed LU
(2027)
In Expert Systems with Applications
334.
p.134414-134414
- Abstract
- This study introduces a reinforcement-learning-based sequential sampling framework for adaptive control charts. Traditional adaptive charts improve monitoring efficiency by adjusting sampling effort according to the current chart state, but they require manually specified decision regions and region-to-action assignments. This becomes difficult when many sample-size and sampling-interval combinations are available. The proposed method preserves the classical EWMA statistic and control-limit signaling rule, while using a deep Q-network (DQN) to learn the adaptive sampling policy for a variable sample size and sampling interval (VSSI) control chart. At each sampling epoch, the agent (DQN) observes the monitoring state and selects the... (More)
- This study introduces a reinforcement-learning-based sequential sampling framework for adaptive control charts. Traditional adaptive charts improve monitoring efficiency by adjusting sampling effort according to the current chart state, but they require manually specified decision regions and region-to-action assignments. This becomes difficult when many sample-size and sampling-interval combinations are available. The proposed method preserves the classical EWMA statistic and control-limit signaling rule, while using a deep Q-network (DQN) to learn the adaptive sampling policy for a variable sample size and sampling interval (VSSI) control chart. At each sampling epoch, the agent (DQN) observes the monitoring state and selects the sampling pair. The general design considers a 70-action space formed from ten candidate sample sizes and seven candidate sampling intervals, with the reward function balancing false-alarm control, detection speed, and sampling-resource constraints. Numerical results show that the DQN-VSSI-EWMA chart satisfies the target average sample size and average sampling interval. The learned policy also exhibits clear VSSI-type behavior, using lower sampling effort in low-risk states and increasing monitoring intensity as the EWMA statistic approaches the control limit. A restricted four-action DQN is also compared with a traditional four-level VSSI-EWMA chart using the same sampling actions. This comparison shows that the DQN can learn a feasible adaptive sampling policy under the same action constraints, while the general 70-action case demonstrates the flexibility of the proposed framework in richer design spaces. The general DQN provides better out-of-control detection for smaller shifts than the restricted four-action DQN, while still satisfying the required in-control constraints. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/9995931f-2384-488e-92ff-4576b23cd6e8
- author
- Sabahno, Hamed
LU
- organization
- publishing date
- 2027-02-01
- type
- Contribution to journal
- publication status
- published
- subject
- in
- Expert Systems with Applications
- volume
- 334
- pages
- 134414 - 134414
- publisher
- Elsevier
- ISSN
- 0957-4174
- DOI
- 10.1016/j.eswa.2026.134414
- language
- English
- LU publication?
- yes
- id
- 9995931f-2384-488e-92ff-4576b23cd6e8
- alternative location
- https://linkinghub.elsevier.com/retrieve/pii/S095741742603318X
- date added to LUP
- 2026-09-16 20:17:03
- date last changed
- 2026-09-17 10:53:57
@article{9995931f-2384-488e-92ff-4576b23cd6e8,
abstract = {{This study introduces a reinforcement-learning-based sequential sampling framework for adaptive control charts. Traditional adaptive charts improve monitoring efficiency by adjusting sampling effort according to the current chart state, but they require manually specified decision regions and region-to-action assignments. This becomes difficult when many sample-size and sampling-interval combinations are available. The proposed method preserves the classical EWMA statistic and control-limit signaling rule, while using a deep Q-network (DQN) to learn the adaptive sampling policy for a variable sample size and sampling interval (VSSI) control chart. At each sampling epoch, the agent (DQN) observes the monitoring state and selects the sampling pair. The general design considers a 70-action space formed from ten candidate sample sizes and seven candidate sampling intervals, with the reward function balancing false-alarm control, detection speed, and sampling-resource constraints. Numerical results show that the DQN-VSSI-EWMA chart satisfies the target average sample size and average sampling interval. The learned policy also exhibits clear VSSI-type behavior, using lower sampling effort in low-risk states and increasing monitoring intensity as the EWMA statistic approaches the control limit. A restricted four-action DQN is also compared with a traditional four-level VSSI-EWMA chart using the same sampling actions. This comparison shows that the DQN can learn a feasible adaptive sampling policy under the same action constraints, while the general 70-action case demonstrates the flexibility of the proposed framework in richer design spaces. The general DQN provides better out-of-control detection for smaller shifts than the restricted four-action DQN, while still satisfying the required in-control constraints.}},
author = {{Sabahno, Hamed}},
issn = {{0957-4174}},
language = {{eng}},
month = {{02}},
pages = {{134414--134414}},
publisher = {{Elsevier}},
series = {{Expert Systems with Applications}},
title = {{A deep reinforcement learning framework for adaptive control charts}},
url = {{http://dx.doi.org/10.1016/j.eswa.2026.134414}},
doi = {{10.1016/j.eswa.2026.134414}},
volume = {{334}},
year = {{2027}},
}