EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
1599. Dynamic operator management in meta-heuristics using machine learning
Invited abstract in session TB-26: Advanced Topics in Combinatorial Optimization, stream Combinatorial Optimization.
Tuesday, 10:30-12:00Room: 012 (building: 208)
Authors (first author is the speaker)
1. | Maryam Karimi Mamaghan
|
Operations Analytics, Vrije Universiteit Amsterdam | |
2. | Mehrdad Mohammadi
|
Industrial Engineering and Innovation Sciences, Eindhoven University of Technology | |
3. | Wout Dullaert
|
Operation analytics, Vrije Universiteit Amsterdam | |
4. | Daniele Vigo
|
DEI, University of Bologna |
Abstract
This study addresses the impact of search operators on the performance of meta-heuristics. The efficient selection and management of operators from a diverse set are essential for optimization. Two key aspects are explored: portfolio selection, determining which subset of operators to consider at each search stage, and operator selection, deciding how to choose operators within that subset for each iteration.
A novel framework is developed to dynamically handle the portfolio of operators, incorporating the tabu concept and Q-learning. Unlike traditional static methods, this study focuses on creating an adaptive online portfolio. Acknowledging that not all operators are useful, the mechanism adds effective operators and excludes ineffective ones at different stages, optimizing overall performance. Inefficient operators are temporarily removed based on a tabu list size and may be reintroduced later. Additionally, new operators may be selectively included for specific search stages. Following portfolio selection, Q-learning dynamically chooses the most efficient operator for each iteration.
To evaluate the proposed framework, the permutation flowshop scheduling problem is tackled. Results demonstrate improvements in optimality gaps and convergence rates compared to offline portfolio selection and existing algorithms. The study highlights the efficacy of the dynamic approach in enhancing meta-heuristic performance and adaptability in complex optimization scenarios.
Keywords
- Metaheuristics
- Machine Learning
- Manufacturing
Status: accepted
Back to the list of papers