Title: Large-State Reinforcement Learning for Hyper-Heuristics

Speaker:Lucas Kletzander
TU Wien, Austria

Summary:
In real-life applications, problems can frequently change or require small adaptations. Manually creating and tuning algorithms for different problem domains or different versions of a problem can be cumbersome and time-consuming. A possible approach is the use of hyper-heuristics, a domain-independent problem solving method where the main task is to select effective chains of problem-specific low-level heuristics on the fly, without knowledge of the problem domain. This task can be seen as a reinforcement learning problem, however, the information available to the hyper-heuristic is very limited.
In this presentation, I will introduce the novel hyper-heuristic LAST-RL (Large-State Reinforcement Learning), using the trajectory of solution changes for a larger set of features, and an exploration strategy based on Iterated Local Search. In addition to the academic HyFlex benchmarks, I will consider several important problems with high practical relevance, which are Rotating Workforce Scheduling, Minimum Shift Design, and Bus Driver Scheduling. Despite being very general, our approach can provide state-of-the-art results on several of these benchmarks, and I will provide insights on how different hyper-heuristics work on our problem domains.

Bio: 
Lucas Kletzander is a PostDoc in the Databases and Artificial Intelligence Group at the Technical University of Vienna, Austria (TU Wien), currently on a research visit at Monash University. He works on complex real-life scheduling problems, many of them directly from industry partners, mainly on employee scheduling with a focus on employee satisfaction aspects, and problems with multiple, potentially conflicting goals. He is interested in a range of exact and heuristic solution methods, with a focus on hybrid methods like Large Neighborhood Search, and problem-independent solution methods like hyper-heuristics. He investigates the use of automated algorithm selection and configuration, as well as machine learning methods in the optimization context. Recent work also deals with interactive methods to guide users through problems with conflicting objectives, and help them understand conflicts and synergies between goals.

—

This event is Hybrid: SEMINAR WILL BEGIN AT 16:00 ON ZOOM 

There will be an in-person meet-and-greet from 15:30 – 16:00 (prior to the seminar):
Melbourne Connect, Level 8, Room 8109, 700 Swanston Street, Carlton 

SEMINAR: WED 9 APRIL 2025 16:00-17:00 (AEST, Melbourne Time)
ZOOM MEETING ID: 873 1557 5255; PASSWORD: 778635