Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/A932B182-EB16-4807-9D8B-776F0426824C | ||
| Year | 2021 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Vasileios Lymperakis, "A novel meta-heuristic search algorithm for global continuous optimization", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2021 https://doi.org/10.26233/heallink.tuc.90358 | ||
| Appears in Collections |
Artificial intelligence research in optimization and search is concerned with reaching the maxima or minima of an objective function, while potentially searching among a vast range of value choices for the function’s variables. Global continuous optimization methods, in particular, seek to reach the optima of complex continuous mathematical functions. Meta-heuristics are commonly used in order to solve such problems. Typically, however, meta-heuristics originally designed for solving discrete optimization problems are later adapted to continuous tasks, which consumes considerable time. Also, there is a chance that they will get stuck to local optima as the complexity of configuration spaces increases. Furthermore, generally meta-heuristics accept worse solutions, in order to achieve a broader exploration of the configuration space. This results in algorithms that do not improve in an anytime manner, and an arbitrary interruption of the algorithm’s flow, can lead to a waste of computation time as the current solution might be worse than a solution discovered earlier on.In this work, a novel single-point meta-heuristic is proposed, which is specifically designed to tackle continuous optimization problems. Our algorithm, Buggy Pinball, is an anytime algorithm inspired by the well-known pinball game: It employs a trajectory-based search, and each proposed solution ensures the improvement of the configuration space. In order to evaluate our algorithm, we used a number of standard testbed functions, which can also be applied on multiple dimensions. We compared our results to the performance of some of the most widely used meta-heuristics, namely simulated annealing, threshold accepting, and particle swarm optimization. Our systematic evaluation shows that our algorithm is very efficient in global continuous optimization tasks, with performance that is particularly successful in complex configuration spaces.
Copyright © DIAS 2013
