7. ESTUDIO DE LA VIABILIDAD TÉCNICA
7.2. CONSUMO DE RECURSOS
Publications and presentations resulting from the pursuit of achieving the re- search aims and objectives defined in Section 1.3.1 are:
• Journal Article Shahin Rostami; Dean O’Reilly; Alex Shenfield; Nick Bowring, “A Novel Preference Articulation Operator for the Evolutionary Multi-Objective Optimisation of Classifiers in Concealed Weapon Detec- tion”, DOI: 10.1016/j.ins.2014.10.031, Volume 295, 20 February 2015, Pages 494520, Information Sciences, Elsevier.
• Conference PaperShahin Rostami; Alex Shenfield, “CMA-PAES: Pareto archived evolution strategy using covariance matrix adaptation for Multi- Objective Optimisation.” Computational Intelligence (UKCI), 2012 12th UK Workshop on. IEEE, 2012.
• Conference Paper Shahin Rostami; Peter Delves; Alex Shenfield, “Evo- lutionary Multi-Objective Optimisation of an Automotive Active Steering Controller”, DOI: 10.13140/2.1.1202.6240 Conference: Manchester Metropoli- tan University Research Symposium 2013.
• Seminar Presentation“Evolutionary Algorithms in Control Systems En- gineering”, 2011 16th November, Seminar, University of Manchester.
• Poster Presentation Shahin Rostami; Alex Shenfield, “Adaptive Grid Archiving Combined with the Covariance Matrix Adaptation Evolution Strategy”, Conference: Manchester Metropolitan University Research Sym- posium 2012.
1.3. Contributions and Objectives 9
The main contributions of this thesis resulting from the pursuit of achieving the research aims and objectives defined in Section 1.3.1 are:
• Development of CMA-PAES, a fast converging EMO algorithm.
This EMO algorithm is inspired by the Pareto Archived Evolution Strategy (PAES) algorithm structure, the AGA method for diversity preservation, and the CMA scheme for variation, with the aim to be light in compu- tational cost, simple in structure, and provide a fast rate of convergence. CMA-PAES has been shown to outperform MO-CMA-ES in this thesis in regards to the quality of the final approximation set paired with the low computational cost of the algorithm overhead. CMA-PAES has been pub- lished in [15] where it is shown to outperform the Nondominated Sorting Genetic Algorithm II (NSGA-II) and PAES, and has been successfully ap- plied to the optimisation of an automotive active steering controller in [16]. A multi-tier variant of CMA-PAES (m-CMA-PAES) is also developed as an EMO algorithm intended for the optimisation of problems containing complex Pareto-optimal sets, by combining a non-elitist AGA based selec- tion scheme with the efficient strategy parameter adaptation of the elitist Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES).
• Development of CMA-PAES-II, a robust many-objective EMO algorithm. This EMO algorithm builds upon the work in CMA-PAES and m-CMA-PAES combined with the new IBC mechanism, sigma restart, and improved AGA. Unlike MO-CMA-ES, CMA-PAES-II allows for the use of CMA in EMO with a computational cost that does not restrict it to
execution on computing clusters or problems consisting of fewer than four objectives.
• Development of the algorithm agnostic1 and novel WZ Preference
Articulation Operator. The operator has the flexibility of being incor- porated a priori, a posteriori or progressively, and as either a primary or auxiliary fitness operator. The two-phase operator has demonstrated the ability to successfully direct the optimisation process closer in proximity to a DM’s expressed ROI, and then proceed to minimise solutions within it. This reduces the computational cost of the optimisation process by re- ducing the scope of the search-space exploration and produces pertinent optimisation sets.
• Incorporation of preference articulation techniques into state of the art EMO algorithms. In this thesis the WZ preference articulation operator has been incorporated into MOEA/D-DRA and CMA-PAES-II, which has shown to improve their performance and the quality in the final approximation set when searching in the presence of DM preferences.
• Successful optimisation of classifiers used for concealed weapon detection. Weighted Z-score Multi-Objective Evolutionary Algorithm Based on Decomposition with Dynamical Resource Allocation (WZ-MOEA/D- DRA) and WZ-CMA-PAES are successfully applied to a real-world opti- misation problem regarding the optimisation of a classifier for concealed 1An operator can be referred to as algorithm agnostic if it has been designed to be easily incorporated into any optimisation framework or algorithm.
1.3. Contributions and Objectives 11
weapon detection, producing better results than previously published clas- sifier implementations. With the confidence instilled from the successful optimisation of the existing solution, new solutions were designed to allow the classification of radar signals into categories of threat objects (e.g. gun, knife, or explosive), which has produced a classifier which would allow for a better response to the detection of a concealed weapon.
Additional contributions that have arisen as a result of this research but are not included in this thesis are:
• Creation of the “EMOLibrary” Evolutionary Multi-Objective Op- timisation Toolbox for MATLAB. The EMOLibrary provides many features that can be utilised in the design of new EMO algorithms, imple- mentation of existing EMO algorithms, or conducting pairwise comparisons of EMO algorithms. The library was inspired by [17] which lacks modern features since it was released in 1994. Features of the EMOLibrary include:
– Performance Metrics, such as the hypervolume indicator, spread, epsilon indicator, generational distance, and inverted generational dis- tance.
– Selection/Sorting Operators, such as non-dominated sorting, the contributing hypervolume indicator, AGA, and the WZ preference ar- ticulation operator.
– Test Problems, such as problems from the following test suites: ZDT, DTLZ, WFG Toolkit, CEC09, and ELLI/CIGTAB test func-
tions. The objective function used for the design of lateral stability controllers (LATCON) for aircraft is also included.
– Parameter Settings, such as problem boundary defaults, suggested reference points, test cases for benchmarking of preference articulation techniques, problem encoding and decoding, problem dimensionality defaults, variable initialisers, and weight generators.
– True Pareto-optimal Fronts for the following test suites: ZDT, DTLZ, WFG Toolkit, and CEC09, with the ability to retrieve the portion of the true Pareto-optimal front within a defined ROI.
• Successful Evolutionary Multi-Objective Optimisation of an Au- tomotive Active Steering Controller. The presented work [16] in- vestigates the use of EMO to optimise the performance of a closed loop feedback Proportional Integral (PI) vehicle yaw controller on a non-linear vehicle. This is done by comparing results against traditional empirical tun- ing methods relating to rise time, settling time, overshoot, and steady-state error. The EMO application showed improvement on the original control tuning and also brought to light the difficulty control engineers face with objective interaction for complex problems.
1.3.1
Research Objectives
With the motivation described in Section 1.1, a number of aims and objectives were defined to guide the direction of work throughout the duration of this re- search. This thesis claims to have achieved every aim and objective in the chapters following. A listing of the aims and objectives are as follows:
1.3. Contributions and Objectives 13
Research Aim
To investigate the incorporation of decision maker preferences into multi-objective evolutionary search methods, so as to improve the quality of final solutions pro- duced by the optimisation process.
Research Objectives
1. To produce a critical review of the field of evolutionary computation with a particular emphasis on using evolutionary computation methods to solve multi-objective problems.
2. To develop and benchmark a state of the art evolutionary multi-objective optimisation algorithm for solving real-world engineering problems, and to provide a basis for incorporating decision maker preferences by enhancing the preservation of diversity across the entire approximation set.
3. To develop a novel algorithm for focussing on regions of interest in multi- objective search spaces.
4. To evaluate the effectiveness of incorporating decision maker preferences into current state-of-the-art evolutionary optimisation routines, using sta- tistically rigorous analysis.
5. To benchmark the new algorithms using synthetic test suites and real-world optimisation problems.