Optimising and automating shift planning
Dandapani, Hariharan (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025072923704
https://urn.fi/URN:NBN:fi:amk-2025072923704
Tiivistelmä
Integral Oy's Intu is an enterprise resource planning system and mobile work media. One of Intu's key features is shift planning, which is currently a manual function in Intu. The objective here was to check the feasibility of automating the process.
However, initial analysis revealed that the actual historical scheduling data from Intu was insufficient, lacking quantifiable employee-related information. Due to these limitations, a carefully curated dataset was synthesised to include major employee attributes, such as certifications, language proficiencies, allergies, and contractual details, while emulating natural working settings. In this sequential synthetic environment, three different approaches were analysed. These were a supervised Random Forest classifier and two reinforcement-based approaches, a Deep Q-Network (DQN) and a Proximal Policy Optimisation (PPO) algorithm. The Random Forest classifier showed very good predictive capability with roughly a 90% success rate for top-1 predictions of employees for the shifts. Reinforcement learning models, especially the model using DQN, exhibited yet another promising performance with about 94% top-1 prediction accuracy. The approach based on PPO followed suit with around 91% top-1 prediction accuracy.
Although the predictive powers of Random Forest and RL-based methods were comparable, reinforcement learning methods have several distinct strengths for real-world practical implementation. One such reinforcement learning benefit is to learn continuously through online training, which can enhance flexibility and responsiveness to dynamic scheduling needs, such as sudden changes in employee availability or unexpected job or shift alterations. Additionally, RL-based models may facilitate the integration of human feedback into training protocols. Such approaches hold the potential for further understanding, refining and individualising model performance, and reducing biases.
However, initial analysis revealed that the actual historical scheduling data from Intu was insufficient, lacking quantifiable employee-related information. Due to these limitations, a carefully curated dataset was synthesised to include major employee attributes, such as certifications, language proficiencies, allergies, and contractual details, while emulating natural working settings. In this sequential synthetic environment, three different approaches were analysed. These were a supervised Random Forest classifier and two reinforcement-based approaches, a Deep Q-Network (DQN) and a Proximal Policy Optimisation (PPO) algorithm. The Random Forest classifier showed very good predictive capability with roughly a 90% success rate for top-1 predictions of employees for the shifts. Reinforcement learning models, especially the model using DQN, exhibited yet another promising performance with about 94% top-1 prediction accuracy. The approach based on PPO followed suit with around 91% top-1 prediction accuracy.
Although the predictive powers of Random Forest and RL-based methods were comparable, reinforcement learning methods have several distinct strengths for real-world practical implementation. One such reinforcement learning benefit is to learn continuously through online training, which can enhance flexibility and responsiveness to dynamic scheduling needs, such as sudden changes in employee availability or unexpected job or shift alterations. Additionally, RL-based models may facilitate the integration of human feedback into training protocols. Such approaches hold the potential for further understanding, refining and individualising model performance, and reducing biases.