Home | Linda's PhD Life
Beranda
Automatic Tuning Parameter Framework
Performance of meta-heuristic algorithm is highly dependent on the parameters value. The right choice of parameter value can have significant effect to the quality of the result. Unfortunately, parameters tuning is not easy and time consuming. As it is mentioned in meta-heuristic literature that in designing and testing new heuristic algorithms, only 10% of the total time is dedicated to algorithm development, while the remaining 90% is used for fine-tuning parameters of the algorithm. The parameters tuning problem will get more difficult when involving different problem instances, where different problem instances may require different parameter configuration.
Motivated by this phenomenon, we started our research with a research question: given a heuristic algorithm and a problem instance, how to set the algorithm parameters so that the algorithm will yield the best possible solution for that instance? To answer this research question, we propose an automatic tuning parameter framework. This framework is consisted of several parts, namely: parameter space reduction, feature-based instances classification, automatic tuning parameter, run-time parameter configuration, run-time feedback analysis and run-time reactive tuning parameter. Currently, we are working on feature-based instances classification and run-time parameter configuration part. The proposed framework is outlined in the picture below.
Contact:
School of Information System
Singapore Management
University
80 Stamford Road
Singapore 178902
Email:
lindawati.2008@phdis.smu.edu.sg