ML to enhance AI Planning for Intelligent Transport Systems
- Reference number
- ID19-0053
- Project leader
- Saffiotti, Alessandro
- Start and end dates
- 200401-240331
- Amount granted
- 2 500 000 SEK
- Administrative organization
- Örebro University
- Research area
- Information, Communication and Systems Technology
Summary
AI Planning and Scheduling (AIPS) methods are fundamental in industrial transport applications, automating key stages in the overall process of assigning tasks and ensuring that plans remain feasible over time. AIPS methods typically rely on manually-specified knowledge to derive plans, many aspects of which are only known to human planning experts (e.g., impenetrable forest roads, icy terrain, etc). AIPS systems should be enhanced with the ability to learn from human planning experts and from experience. The use of learning will allow AIPS systems to scale up and provide good quality plans in many transport domains. The four-year project will involve an Industrial PhD student, employed at Scania, and co-supervised by the main applicant at Örebro University, and will lead to a Doctorate degree. Three tasks are envisaged, focusing on (1) how to represent and reason about knowledge that is learned from execution traces, (2) learning knowledge from human experts and from experience in simulation, and (3) exploiting learned knowledge to achieve robustness and real-world applicability. The project is expected to advance the state of the art in automated planning, significantly pushing forward the applicability of such methods to real-world problems. The wide applicability of the achieved results amplifies the impact of the developed technology, which will increase economic and ecological sustainability in a vast range of industries.
Popular science description
AI planning and scheduling studies the problem of automatically deriving plans. A plan is a choice and temporal organization of tasks that should be carried out in order to achieve as best as possible some pre-stated objectives. This problem lies at the core of Artificial Intelligence and has been studied for decades. However, existing methods are difficult to apply in real-world applications, and in particular, those involving transport systems. This project will study methods for improving the applicability of AI planning and scheduling methods via machine learning. The key insight is to mimic what human planners do, that is, learn from positive and negative examples, and from more knowledgeable humans. It is expected that the use of machine learning will allow automated planning and scheduling systems to scale up and provide good quality plans in many domains. This project will target in particular industrial transport domains. Due to the ubiquity of such application contexts, we expect the results of the project to significantly increase the economic and ecological sustainability of a vast range of industries in which transport needs are pivotal. These include mining, logistics, and public transport, only to name a few.