Offer 1
How Far Can Deep Reinforcement Learning Scale for Operational Planning?
English summary
This IDP investigates how the scalability of deep reinforcement learning-based operational planning changes with problem size. Students will implement a DRL training framework for comparable CPU- and GPU-based experiments, vary problem and system parameters, measure computational and performance metrics, and evaluate bottlenecks and GPU acceleration.
Overview
- Project details
- How Far Can Deep Reinforcement Learning Scale for Operational Planning?
Source excerpt · Page 1
- “How Far Can Deep Reinforcement Learning Scale for Operational Planning?” Page 1
- Organization
- Not stated in the PDF
- Project formats
- IDP
Source excerpt · Page 1
- “IDP” Page 1
- IDP
- Degree levels
- Master
Source excerpt · Page 1
- “The thesis is suitable for master’s students in Computer Science.” Page 1
- Master
- Project goal
- Investigate the scalability boundary of current DRL-based planning methods.
Source excerpt · Page 1
- “The project will analyze how training time, memory usage, sample efficiency, and policy quality change as the size of the operational planning problem increases.” Page 1
Topics and work
- Subjects
- Deep reinforcement learning (DRL)
Source excerpt · Page 1
- “Deep Reinforcement Learning (DRL) has shown strong potential for solving complex sequential decision-making problems in operations management” Page 1
- operational planning
Source excerpt · Page 1
- “realistic, large-scale operational planning problems” Page 1
- scalability of DRL-based planning methods
Source excerpt · Page 1
- “The objective of this thesis is to investigate the scalability boundary of current DRL-based planning methods.” Page 1
- GPU-based simulation and policy training
Source excerpt · Page 1
- “A particular focus will be placed on understanding whether GPU-based simulation and policy training can provide the speedups required for practically relevant large-scale applications.” Page 1
- Deep reinforcement learning (DRL)
- Application areas
- inventory control
Source excerpt · Page 1
- “including inventory control, production planning, routing, scheduling, and resource allocation.” Page 1
- production planning
Source excerpt · Page 1
- “including inventory control, production planning, routing, scheduling, and resource allocation.” Page 1
- routing
Source excerpt · Page 1
- “including inventory control, production planning, routing, scheduling, and resource allocation.” Page 1
- scheduling
Source excerpt · Page 1
- “including inventory control, production planning, routing, scheduling, and resource allocation.” Page 1
- resource allocation
Source excerpt · Page 1
- “including inventory control, production planning, routing, scheduling, and resource allocation.” Page 1
- inventory control
- Methods and tools
- Deep Reinforcement Learning (DRL)
Source excerpt · Page 1
- “Deep Reinforcement Learning (DRL) has shown strong potential for solving complex sequential decision-making problems in operations management” Page 1
- CPU- and GPU-based experiments
Source excerpt · Page 1
- “implement a DRL training framework that can run comparable CPU- and GPU-based experiments.” Page 1
- GPU-based simulation
Source excerpt · Page 1
- “whether GPU-based simulation and policy training can provide the speedups required for practically relevant large-scale applications.” Page 1
- policy training
Source excerpt · Page 1
- “whether GPU-based simulation and policy training can provide the speedups required for practically relevant large-scale applications.” Page 1
- Deep Reinforcement Learning (DRL)
- Activities
- Define scalable operational planning problems
Source excerpt · Page 1
- “Define scalable operational planning problems” Page 1
- Implement a DRL training framework for comparable CPU- and GPU-based experiments
Source excerpt · Page 1
- “implement a DRL training framework that can run comparable CPU- and GPU-based experiments.” Page 1
- Systematically vary problem size, batch size, network size, and simulator design
Source excerpt · Page 1
- “Systematically vary problem size, batch size, network size, and simulator design” Page 1
- Measure training time, memory usage, sample throughput, and performance
Source excerpt · Page 1
- “to measure training time, memory usage, sample throughput, and performance.” Page 1
- Identify computational bottlenecks
Source excerpt · Page 1
- “Identify the main computational bottlenecks” Page 1
- Evaluate whether GPU acceleration enables practically relevant scaling
Source excerpt · Page 1
- “evaluate whether GPU acceleration enables practically relevant scaling.” Page 1
- Define scalable operational planning problems
- Kinds of work
- Expected outputs
- DRL training framework
Source excerpt · Page 1
- “implement a DRL training framework that can run comparable CPU- and GPU-based experiments.” Page 1
- DRL training framework
Requirements
- Required skills
- Ability to work independently
Source excerpt · Page 1
- “The ability to work independently, as well as analytical skills, are required.” Page 1
- Analytical skills
Source excerpt · Page 1
- “The ability to work independently, as well as analytical skills, are required.” Page 1
- Profound knowledge of Python
Source excerpt · Page 1
- “Profound knowledge of Python (preferably JAX / PyTorch) is required.” Page 1
- Ability to work independently
- Recommended skills
- Eligible study fields
- Computer Science
Source excerpt · Page 1
- “The thesis is suitable for master’s students in Computer Science.” Page 1
- Computer Science
- Programming in the project
- Central
Source excerpt · Page 1
- “implement a DRL training framework that can run comparable CPU- and GPU-based experiments.” Page 1
- Programming prerequisite
- Required
Source excerpt · Page 1
- “Profound knowledge of Python (preferably JAX / PyTorch) is required.” Page 1
Practical details
- Team size
- Not stated in the PDF
- Duration
- Not stated in the PDF
- Start
- October 2026Normalized date: 2026-10
Source excerpt · Page 1
- “Earliest begin: October 2026” Page 1
- Application deadline
- Not stated in the PDF
- Work location
- Not stated in the PDF
- Location mode
- Not stated in the PDF
- Working language
- Not stated in the PDF
Learning and support
- Learning opportunities
- Not stated in the PDF
- Support offered
- Not stated in the PDF
Application and contacts
- Contacts
- Patrick Helm
Source excerpt · Page 1
- “Supervisor: Patrick Helm” Page 1
- logtheses.log@mgt.tum.de
Source excerpt · Page 1
- “Email with curriculum vitae and transcript of records to logtheses.log@mgt.tum.de” Page 1
- Patrick Helm
- How to apply
- Email with curriculum vitae and transcript of records to logtheses.log@mgt.tum.de
Source excerpt · Page 1
- “Email with curriculum vitae and transcript of records to logtheses.log@mgt.tum.de” Page 1
- Further information
- Not stated in the PDF