My PhD Thesis

I've finally got around to uploading a close-to-final draft of my PhD thesis (defended in early 2015). As the title suggests, the work is focused on describing several enhancements that can help multi-objective evolutionary algorithms (MOEAs) solve computationally-intensive optimization problems in a decent time frame. Large parts of the main novel scientific contributions of this thesis have been disseminated a priori in 7 peer-reviewed articles (3 journals and 4 conferences and edited volumes).

The practical scenarios I've tackled are all related to multi-objective optimizations of electrical drive designs and the very good results I've obtained on them also stem from a close  and very fruitful cooperation with the Linz Center of Mechatronics (LCM) and the Institute of Electrical Drives and Power Electronics (EAL) of the Johannes Kepler University Linz. Hopefully the version I've uploaded doesn't contain too many mistakes (scientific or language wise), but if you find any, please do let me know.

You can download a close-to-final draft of the paper by clicking here or from my Downloads box (Zavoianu Ciprian - PhD Thesis.pdf). The same version of the document can be previewed at the bottom of this post. For the official, final version of the document, please refer to the Johannes Kepler University Library. If you have any questions regarding the work, please don't hesitate to contact me.

For citations please use the following BibTeX reference:

@PhdThesis{Zavoianu2015PhD,

  author = {Alexandru-Ciprian Z\u{a}voianu},

  title = {Enhanced Evolutionary Algorithms for Solving Computationally-Intensive Multi-Objective Optimization Problems},

  school = {Johannes Kepler University Linz, Austria},

  year = {2015},

  month = {January}

}

Paper: Performance Comparison of Generational and Steady-State Asynchronous Multi-Objective Evolutionary Algorithms for Computationally-Intensive Problems

This journal article describes findings in one of the three main research foci I've investigated during my PhD research project. It is an improved and extended version of the conference manuscript prepared for ICAISC 2013. Here is the abstract of the article:

In the last two decades, multi-objective evolutionary algorithms (MOEAs) have become ever more used in scientific and industrial decision support and decision making contexts the require an a posteriori articulation of preference. The present work is focused on a comparative analysis of the performance of two master-slave parallelization (MSP) methods, the canonical generational scheme and the steady-state asynchronous scheme. Both can be used to improve the convergence speed of multi-objective evolutionary algorithms that must use computationally-intensive fitness evaluation functions. Both previous and present experiments show that a correct choice for one or the other parallelization method can lead to substantial improvements with regard to the overall duration of the optimization process. Our main aim is to provide practitioners of MOEAs with a simple but effective method of deciding which MSP option is better given the particularities of the concrete optimization process. This in turn, would give the decision maker more time for articulating preferences (i.e., more flexibility). Our analysis is performed based on 15 well-known MOOP benchmark problems and two simulation-based industrial optimization process from the field of electrical drive design. For the first industrial MOOP, when comparing with a preliminary study, applying the steady state asynchronous MSP enables us to achieve an overall speedup (in terms of total wall-clock computation time) of ≈ 25%. For the second industrial MOOP, applying the steady-state MSP produces an improvement of ≈ 12%. We focus our study on two of the best known and most widely used MOEAs: the Non-dominated Sorting Genetic Algorithm II (NSGA-II) and the Strength Pareto Evolutionary Algorithm (SPEA2).

You can download the preprint version of the paper by clicking here or from my Downloads box (Performance Comparison of Generational and Steady-State Asynchronous Multi-Objective Evolutionary Algorithms for Computationally-Intensive Problems - KBS 2015.pdf). The same preprint version can be previewed at the bottom of this post. The final publication is available at elsevier.com.

For citations please use the following BibTeX reference:

@ARTICLE{Zavoianu2015KBS,
  author = {Alexandru-Ciprian Z\u{a}voianu and Edwin Lughofer and Werner Koppelst\"{a}tter and G\"{u}nther Weidenholzer and Wolfgang Amrhein and Erich Peter Klement},
  title = {Performance comparison of generational and steady-state asynchronous multi-objective evolutionary algorithms for computationally-intensive problems},
  journal = {Knowledge-Based Systems},

  note = {Computational Intelligence Applications for Data Science},
  year = {2015},
  volume = {87},
  pages = {47-60},
  doi = {10.1016/j.knosys.2015.05.029}
}

Paper: DECMO2 - A Robust Hybrid and Adaptive Multi-Objective Evolutionary Algorithm

This journal article also stems from my PhD research project - enhancing currently available Evolutionary Computation methods employed for solving computationally-intensive multi-objective optimization problems. The presented algorithm - DECMO2 - is an improved version of a method presented in one of our earlier papers. Here is the abstract of the article:

We describe a hybrid and adaptive coevolutionary optimization method that can efficiently solve a wide range of multi-objective optimization problems (MOOPs) as it successfully combines positive traits from three main classes of multi-objective evolutionary algorithms (MOEAs): classical approaches that use Pareto-based selection for survival criteria, approaches that rely on differential evolution, and decomposition-based strategies. A key part of our hybrid evolutionary approach lies in the proposed fitness sharing mechanism that is able to smoothly transfer information between the coevolved subpopulations without negatively impacting the specific evolutionary process behavior that characterizes each subpopulation. The proposed MOEA also features an adaptive allocation of fitness evaluations between the coevolved populations in order to increase robustness and favor the evolutionary search strategy that proves more successful for solving the MOOP at hand. Apart from the new evolutionary algorithm, this paper also contains the description of a new hypervolume and racing-based methodology aimed at providing practitioners from the field of multi-objective optimization with a simple means of analyzing/reporting the general comparative run-time performance of multi-objective optimization algorithms over large problem sets.

You can download the preprint version of the paper by clicking here or from my Downloads box (DECMO2 - A Robust Hybrid and Adaptive Multi-Objective Evolutionary Algorithm - SOCO 2014.pdf). The same preprint version can be previewed at the bottom of this post. The final publication is available at Springer via http://dx.doi.org/10.1007/s00500-014-1308-7.

For citations please use the following BibTeX reference:

@ARTICLE{Zavoianu2014SOCO,
  author = {Alexandru-Ciprian Z\u{a}voianu and Edwin Lughofer and Gerd Bramerdorfer and Wolfgang Amrhein and Erich Peter Klement},
  title = {{DECMO2}: a robust hybrid and adaptive multi-objective evolutionary algorithm},
  journal = {Soft Computing},
  year = {2014},
  volume = {19},
  number = {12},
  pages = {3551-3569},
  doi = {10.1007/s00500-014-1308-7}
}

Paper: An Effective Ensemble-Based Method for Creating On-the-Fly Surrogate Fitness Functions for Multi-Objective Evolutionary Algorithms - SYNASC 2013

This paper was prepared for the SYNASC 2013 conference and it is related to my current research project that has the general aim of enhancing currently available Evolutionary Computation methods employed for the multi-objective optimization of problems that rely on a very time-intensive fitness evaluation functions. Here is the abstract of the article:

The task of designing electrical drives is a multi-objective optimization problem (MOOP) that remains very slow even when using state-of-the-art approaches like particle swarm optimization and evolutionary algorithms because the fitness function used to assess the quality of a proposed design is based on time-intensive finite element (FE) simulations. One straightforward solution is to replace the original FE-based fitness function with a much faster-to-evaluate surrogate. In our particular case each optimization scenario poses rather unique challenges (i.e., goals and constraints) and the surrogate models need to be constructed on-the-fly, automatically, during the run of the evolutionary algorithm. In the present research, using three industrial MOOPs, we investigated several approaches for creating such surrogate models and discovered that a strategy that uses ensembles of multi-layer perceptron neural networks and Pareto-trimmed training sets is able to produce very high-quality surrogate models in a relatively short time interval.

You can download a preliminary version of the paper by clicking here or from my Downloads box (On the Performance of Master-Slave Parallelization Methods for MOEAs - ICAISC 2013.pdf). The same preliminary draft of the document can be previewed at the bottom of this post. The original publication is available at the IEEE Xplore website.

For citations please use the following BibTeX reference:

@INPROCEEDINGS{Zuavoianu2013SYNASC,
  author = {Alexandru-Ciprian Z\u{a}voianu and Edwin Lughofer and Gerd Bramerdorfer and Wolfgang Amrhein and Erich Peter Klement},
  title = {An Effective Ensemble-Based Method for Creating On-the-Fly Surrogate Fitness Functions for Multi-Objective Evolutionary Algorithms},
  booktitle = {Proceedings of the 15th International Symposium on Symbolic and Numeric Algorithms for Scientific Computing (SYNASC 2013)},
  year = {2013},
  pages = {235-248},
  publisher = {IEEE Computer Society},
}

Paper: Efficient Multi-Objective Optimization Using 2-Population Cooperative Coevolution - EUROCAST 2013

This paper was prepared for the EUROCAST 2013 conference and contains a proof-of-concept for DECMO - a simple coevolutionary multi-objective optimization algorithm that, on several problems is successfully able to adopt the search behavior of the most successful strategy it incorporates. This line of research was continued and fully fleshed out in DECMO2 - a subsequent algorithm I developed for my PhD. Here is the abstract our EUROCAST 2013 submission:

We propose a 2-population cooperative coevolutionary optimization method that can efficiently solve multi-objective optimization problems as it successfully combines positive traits from classic multi-objective evolutionary algorithms and from newer optimization approaches that explore the concept of differential evolution. A key part of the algorithm lies in the proposed dual fitness sharing mechanism that is able to smoothly transfer information between the two coevolved populations without negatively impacting the independent evolutionary process behavior that characterizes each population.

You can download a preliminary version of the paper by clicking here or from my Downloads box (Efficient Multi-Objective Optimization Using 2-Population Cooperative Coevolution - EUROCAST 2013.pdf). The same preliminary draft of the document can be previewed at the bottom of this post. The original publication is available at www.springerlink.com.

For citations please use the following BibTeX reference:

@INCOLLECTION{Alexandru-CiprianZavoianu2013,
  author = {Alexandru-Ciprian Z\u{a}voianu and Edwin Lughofer and Wolfgang Amrhein and Erich Peter Klement},
  title = {Efficient Multi-Objective Optimization Using 2-Population Cooperative Coevolution},
  booktitle = {Computer Aided Systems Theory - EUROCAST 2013},
  publisher = {Springer Berlin / Heidelberg},
  year = {2013},
  editor = {Moreno-Diaz, Roberto and Pichler, Franz and Quesada-Arencibia, Alexis},
  volume = {8111},
  series = {Lecture Notes in Computer Science},
  pages = {251-258},
  doi = {10.1007/978-3-642-53856-8_32},
}

Paper: Hybridization of Multi-Objective Evolutionary Algorithms and Artificial Neural Networks for Optimizing the Performance of Electrical Drives - EAAI

A big part of my current PhD work at the JKU-Department of Knowledge-Based Mathematical Systems is related to the dissemination of our current research (i.e., writing scientific articles). That's why most of my recent / forthcoming posts are / will be about "papers".

The work presented in this post is a revised and extended (journal) version of one of the earlier papers written in collaboration with our partners from the Institute for Electrical Drives and Power Electronics of the Johannes Kepler University, Linz. The aim of the article is to describe a surrogate-based enhancement that can help to significantly speed-up a multi-objective evolutionary algorithm that requires an extremely time-intensive fitness evaluation function. Here is the abstract of the article:

Performance optimization of electrical drives implies a lot of degrees of freedom in the variation of design parameters, which in turn makes the process overly complex and sometimes impossible to handle for classical analytical optimization approaches. This, and the fact that multiple non-independent design parameter have to be optimized synchronously, makes a soft computing approach based on multi-objective evolutionary algorithms (MOEAs) a feasible alternative. In this paper, we describe the application of the well known Non-dominated Sorting Genetic Algorithm II (NSGA-II) in order to obtain high-quality Pareto-optimal solutions for three optimization scenarios. The nature of these scenarios requires the usage of fitness evaluation functions that rely on very time-intensive finite element (FE) simulations. The key and novel aspect of our optimization procedure is the on-the-fly automated creation of highly accurate and stable surrogate fitness functions based on artificial neural networks (ANNs). We employ these surrogate fitness functions in the middle and end parts of the NSGA-II run (=> hybridization) in order to significantly reduce the very high computational effort required by the optimization process. The results show that by using this hybrid optimization procedure, the computation time of a single optimization run can be reduced by 46% to 72% while achieving Pareto-optimal solution sets with similar, or even slightly better, quality as those obtained when conducting NSGA-II runs that use FE simulations over the whole run-time of the optimization process.

You can download the preprint version of the paper by clicking here or from my Downloads box (Hybridization of Multi-Objective Evolutionary Algorithms and Artificial Neural Networks for Optimizing the Performance of Electrical Drives - EAAI 2013.pdf). The same preprint version can be previewed at the bottom of this post. The original publication is available at elsevier.com.

For citations please use the following BibTeX reference:

@ARTICLE{Zavoianu2013EAAI,
  author = {Alexandru-Ciprian Z\u{a}voianu and Gerd Bramerdorfer and Edwin Lughofer and Siegfried Silber and Wolfgang Amrhein and Erich Peter Klement},
  title = {Hybridization of Multi-Objective Evolutionary Algorithms and Artificial Neural Networks for Optimizing the Performance of Electrical Drives},
  journal = {Engineering Applications of Artificial Intelligence},
  year = {2013},
  volume = {26},
  pages = {1781-1794},
  number = {8},
  doi = {10.1016/j.engappai.2013.06.002}
}

Paper: On the Performance of Master-Slave Parallelization Methods for Multi-Objective Evolutionary Algorithms - ICAISC 2013

This paper was prepared for the ICAISC 2013 conference and it is related to my current research project that has the general aim of enhancing currently available Evolutionary Computation methods employed for the multi-objective optimization of problems that rely on a very time-intensive fitness evaluation functions. Here is the abstract of the article:

This paper is focused on a comparative analysis of the performance of two master-slave parallelization methods, the basic generational scheme and the steady-state asynchronous scheme. Both can be used to improve the convergence speed of multi-objective evolutionary algorithms (MOEAs) that rely on time-intensive fitness evaluation functions. The importance of this work stems from the fact that a correct choice for one or the other parallelization method can lead to considerable speed improvements with regards to the overall duration of the optimization. Our main aim is to provide practitioners of MOEAs with a simple but effective method of deciding which master-slave parallelization option is better when dealing with a time-constrained optimization process.

You can download a preliminary version of the paper by clicking here or from my Downloads box (On the Performance of Master-Slave Parallelization Methods for MOEAs - ICAISC 2013.pdf). The same preliminary draft of the document can be previewed at the bottom of this post. The original publication is available at www.springerlink.com.

For citations please use the following BibTeX reference:

@INCOLLECTION{Zavoianu2013ICAISC,
  author = {Alexandru-Ciprian Z\u{a}voianu and Edwin Lughofer and Werner Koppelst\"{a}tter and G\"{u}nther Weidenholzer and Wolfgang Amrhein and Erich Peter Klement},
  title = {On the Performance of Master-Slave Parallelization Methods for Multi-Objective Evolutionary Algorithms},
  booktitle = {Artificial Intelligence and Soft Computing},
  publisher = {Springer Berlin Heidelberg},
  year = {2013},
  editor = {Laszek Rutkowski et al.},
  volume = {7895},
  series = {Lecture Notes in Artificial Intelligence},
  pages = {122-134},
  doi = {10.1007/978-3-642-38610-7_12}
}

Paper: A Hybrid Soft Computing Approach for Optimizing Design Parameters of Electrical Drives - SOCO 2012

This paper was prepared for the SOCO 2012 conference and it provides some insight into my current research project that has the general aim of enhancing currently available Evolutionary Computation methods employed for the optimization of design parameters used in the prototyping process of electrical drives. The present paper is particularly related to this earlier post regarding the treatment of expensive fitness evaluation functions in Evolutionary Computation. Here is the abstract of the article:

In this paper, we are applying a hybrid soft computing approach for optimizing the performance of electrical drives where many degrees of freedom are allowed in the variation of design parameters. The hybrid nature of our approach originates from the application of multi-objective evolutionary algorithms (MOEAs) to solve the complex optimization problems combined with the integration of non-linear mappings between design and target parameters. These mappings are based on artificial neural networks (ANNs) and they are used for the fitness evaluation of individuals (design parameter vectors). The mappings substitute very time-intensive finite element simulations during a large part of the optimization run. Empirical results show that this approach finally reduces the computation time for single runs from a few days to several hours while achieving Pareto fronts with a similar high quality.

You can download a preliminary version of the paper by clicking here or from my Downloads box (A Hybrid Soft Computing Approach for Optimizing Design Parameters of Electrical Drives - SOCO 2012.pdf). The same preliminary draft of the document can be previewed at the bottom of this post. The original publication is available at www.springerlink.com.

For citations please use the following BibTeX reference:

@INCOLLECTION{ Alexandru-CiprianZavoianu2013,
  author = {Alexandru-Ciprian Z\u{a}voianu and Gerd Bramerdorfer and Edwin Lughofer and Siegfried Silber and Wolfgang Amrhein and Erich Peter Klement},
  title = {A Hybrid Soft Computing Approach for Optimizing Design Parameters of Electrical Drives},
  booktitle = {Soft Computing Models in Industrial and Environmental Applications},
  publisher = {Springer Berlin Heidelberg},
  year = {2013},
  editor = {V\'{a}clav Sn\'{a}\u{s}el and Ajith Abraham and Emilio S. Corchado},
  volume = {188},
  series = {Advances in Intelligent Systems and Computing},
  pages = {347-358},
  doi = {10.1007/978-3-642-32922-7_36}
}

Paper: Improving the Parsimony of Regression Models for an Enhanced Genetic Programming Process - EUROCAST 2011

This paper was prepared for the EUROCAST 2011 conference and it basically contains a short summary of the work I have carried out for my Master Thesis Project. Here is the abstract of the article:

This research is focused on reducing the average size of the solutions generated by an enhanced GP process without affecting the high predictive accuracy this method exhibits when being applied on a complex, industry proposed, regression problem. As such, the effects the GP enhancements have on bloat have been studied and, finally, a bloat control system based on dynamic depth limiting (DDL) and iterated tournament pruning (ITP) was designed. The resulting bloat control system is able to improve by approx. 40% the average GP solution parsimony without impacting average solution accuracy.

You can download a preliminary version of the paper by clicking here or from my Downloads box (Improving the Parsimony of Regression Models for an Enhanced GP Process - EUROCAST 2011.pdf). The same preliminary draft of the document can be previewed at the bottom of this post. The original publication is available at www.springerlink.com.

For citations please use the following BibTeX reference:

@INCOLLECTION{Alexandru-CiprianZavoianu2012,
  author = {Alexandru-Ciprian Z\u{a}voianu and Gabriel Kronberger and Michael Kommenda and Daniela Zaharie and Michael Affenzeller},
  title = {Improving the Parsimony of Regression Models for an Enhanced Genetic Programming Process},
  booktitle = {Computer Aided Systems Theory - EUROCAST 2011},
  publisher = {Springer Berlin / Heidelberg},
  year = {2012},
  editor = {Moreno-Diaz, Roberto and Pichler, Franz and Quesada-Arencibia, Alexis},
  volume = {6927},
  series = {Lecture Notes in Computer Science},
  pages = {264-271},
  doi = {10.1007/978-3-642-27549-4_34},
}

My Master's Thesis

As promised in my previous post regarding code bloating in GP, here is my Master's Thesis. Hopefully the version I've uploaded doesn't contain too many mistakes (scientific or language wise), but if you find any, please do let me know.

You can download a close-to-final draft of the paper by clicking here or from my Downloads box (Zavoianu Ciprian - Master's Thesis.pdf). The same version of the document can be previewed at the bottom of this post. For the official, final version of the document, please refer to the Johannes Kepler University Library. If you have any questions regarding it, please don't hesitate to contact me.

For citations please use the following BibTeX reference or refer to the conference article that summarizes this work:

@MASTERSTHESIS{Zavoianu2010,

  author = {Alexandru-Ciprian Z\u{a}voianu},

  title = {Towards solution parsimony in an enhanced genetic 

    programming process},

  school = {International School Informatics: Engineering \& 

    Management, ISI-Hagenberg, Johannes Kepler University, Linz},

  year = {2010}

}

Paper: On Sensor Data Simulation - SYNASC 2009

I spent most of my time at the e-Austria Research Institute in Timişoara, Romania working with a team that was involved in the European FP7 Research Project DEHEMS. My main task was to design and implement a robust sensor simulation system. I worked on this task with my colleague, Ciprian Craciun, who's insight into the matter at hand and experience helped tremendously.
.

The resulting simulation system (coded in Java) was fully described in a paper prepared for the SYNASC 2009 conference. Here is the abstract of the paper:

In the context of the DEHEMS project, aiming to analyze real time sensor readings related with the energy performance of individual hoseholds, the research and development stages of various modules have underlined the need of a fast and scalable means for generating large amounts of high quality sensor data. This paper introduces a multi-purpose sensor simulation system (DS3 ) that successfully meets all the data generation requirements imposed by DEHEMS . Primarily designed as an internal tool for testing and tuning the storage and data analysis engines, DS3 will act also as a support platform for a number of end-user orientated applications.

You can download a preliminary version of the paper by clicking here or from my Downloads box (On Sensor Data Simulation - SYNASC 2009.pdf). The same preliminary draft of the document can be previewed at the bottom of this post. The original publication is available at the IEEE Xplore website.

For citations please use the following BibTeX reference:

@INPROCEEDINGS{Alexandru-CiprianZavoianu2009,

  author = {Alexandru-Ciprian Z\u{a}voianu and Ciprian-Dorin Craciun},
  title = {On Sensor Data Simulation},
  booktitle = {Proceedings of the 11th International Symposium on Symbolic and Numeric Algorithms for Scientific Computing (SYNASC 2009)},
  year = {2009},
  pages = {420-425},
  publisher = {IEEE Computer Press},
  doi = {10.1109/SYNASC.2009.29}
}

My Bachelor's Thesis

I've finally got around to converting my Bachelor's Thesis to .PDF and uploading it. Unfortunately, it is entirely in Romanian. As I've previously stated its main focus is to describe a hybrid lightweight task scheduling algorithm based on metaheuristics like local search, tabu search and simulated annealing (an adaptation of the proposed algorithm for the university course timetabling problem is also described). My research for this paperwas coordinated by Prof. Dr. Dana Petcu. Here is the abstract:

This paper presents a general purpose hybrid timetabling algorithm based on metaheuristics and describes an implementation which is specifically aimed at solving the university course timetabling problem.
The first chapter presents some introductory information regarding the problem at hand, the current state of knowledge and briefly describes the timetabling method we propose. The second chapter describes with great detail the local search techniques we rely on and the base algorithm our method was inspired by: the CBRS algorithm. This part of the paper also contains the detailed description of our timetabling algorithm, justifications regarding the need of a multi-threading approach and of a thread trimming mechanism, together with a pseudo-code representation.
The third chapter describes the implementation of the algorithm in a Borland Delphi component and the last section of this paper presents some test results obtained by a timetabling application that incorporates our Delphi component and the final conclusions regarding the effectiveness of both the general timetabling method we have proposed and of its implementation.

You can download the paper by clicking here or from my Downloads box (Zavoianu Ciprian - Lucrare de diploma.pdf). If you have any questions regarding it, please don't hesitate to contact me.