Master's thesis · Bosch
Can machine learning estimate task durations?
Built four prediction models on real Bosch schedule data. The lesson: fix the data before the AI.
At a glance
- Type
- Master's thesis, M.Eng. Technology & Innovation Management
- Period
- Oct 2023 – May 2024
- Data
- 4,000+ tasks from two Bosch projects
- Tools
- MS Project, Python (pandas, scikit-learn), MySQL, Signavio, Visio
About
Written in the Project Management Office of Bosch Semiconductor Operations, Reutlingen, as my master's thesis at Hochschule Harz.
The challenge
In large projects, the duration of each task is still mostly estimated by expert judgement. The PMO wanted to know whether artificial intelligence could help in the planning phase, and which kind of AI would actually be useful.
The setting
The thesis was written at Bosch’s wafer fab in Reutlingen, the company’s most established semiconductor site, making chips since 1970. Reutlingen produces automotive ICs, MEMS sensors (the tiny motion and pressure sensors in airbags, ESC and smartphones) and power semiconductors, including silicon carbide MOSFETs, all on 200 mm wafers. It is the only Bosch site that combines frontend and backend manufacturing with its own test centre.
The data came from two large international projects of the PMO.
What I did
- Reviewed the research: a systematic literature search of 200+ sources, narrowed down to 40+, on AI in schedule, risk and resource management.
- Analysed how planning really works: interviews with experts, direct observation, and a comparison of the PMI standard, the Bosch standard and a live project.
- Built a data pipeline: exported the project schedules from MS Project, cleaned and encoded them, and loaded them into a MySQL database.
- Developed and trained four models in Python: Random Forest, Gradient Boosting, a neural network and support vector regression.
Result
The two ensemble models clearly won. The best model, a Random Forest, explained over 85% of the variation in task durations. The neural network and support vector regression fell well short with this amount of data. I presented the results to Bosch management after the defence.
What I learned
Most research recommends neural networks, but with real company data the quality and amount of data matter more than the algorithm. Before jumping on the AI bandwagon, fix the data foundation. Next steps I proposed: improve the data set, add resources as a feature, connect data between planning processes and include real-time factors for more resilient schedules.