Interest in the integration of evolutionary computing (EC) with engineering design has rapidly increased over the past ten years as practical application illustrating their potential when utilized as design search, exploration and optimization techniques has become more evident. Changes in the manner in which we practise design have also greatly facilitated the successful integration of EC with day-to-day activities. Powerful inexpensive desktop computing capability has become increasingly available and this, coupled with more usable and accessible modelling environments and analysis codes within web/grid-based engineering problem-solving environments increasingly supports the designer in the coupling of tools and solvers to address specific problem areas. Evolutionary algorithms are now playing a significant role within suites of optimization software available to the engineer and their integration with routine design tasks can now be implemented relatively easily. Several commercial optimization software packages now include evolutionary algorithms, and packages solely exploiting EC techniques are becoming available.
In addition to the core evolutionary computing algorithms such as evolution strategies, genetic algorithms, evolutionary programming and genetic programming, other adaptive search algorithms that possess many, if not all, of the performance characteristics of the evolutionary algorithms are now considered to part of the EC field. These include ant colony models, tabu search, scatter search and simulated annealing. These techniques have become firmly established and are now more widely utilized within industry. Most of these algorithmic search strategies are, in the main, analogous to natural systems and depend to a greater or lesser extent upon agent interaction and information exchange with search resource allocation dependant upon the relative fitness of identified solutions. The main evolutionary algorithms adopt a population-based, generational approach to search and optimization. That is, the process generally commences from a population of solutions each comprising randomly generated values of those variable parameters that describe the system to be optimized. The relative fitness of each solution can then be calculated from some form of computational model of the system under design and it is this relative fitness which generally determines whether a solution will survive into the next generation. Selection is a stochastic process where those solutions of relatively high fitness have a greater probability of survival and duplication. Reproduced individuals then become the parents of the next generation. They are stochastically paired and a recombination operator is introduced that randomly selects material from each parent in order to create offspring. These offspring then represent the members of the new generation. A random mutation operator is also introduced to perform small numbers of random perturbations to variable values within the overall population. This supports diversity and exploration whilst preventing premature convergence of the system.
The common attributes of the various stochastic search techniques of particular relevance to engineering design processes include a requirement for little, if any, a priori knowledge relating to the search environment as no gradient information is needed. EC techniques can therefore successfully negotiate search spaces described by a wide range of model types and structure. The techniques possess excellent exploratory capabilities especially where population-based search is considered. Such techniques initially randomly populate a design space with trial solutions and the extent of subsequent search from initial points depends upon their relative performance. Further sampling of diverse regions of that space can continue throughout the search process through the action of stochastic mutation operators. The stochastic nature of the various algorithms combined with continuing random sampling of the search space can prevent convergence upon local sub-optima. The algorithms generally have the ability to handle high dimensionality and are robust across a wide range of problem classes. They can identify multiple high-performance solutions from complex design spaces and multi-objective techniques such as Pareto can also be easily and successfully integrated to provide a range of good solutions for further off-line evaluation.
Current state-of-the-art now provides an indication of the best way forward to ensure short- to medium-term industrial take-up of the technologies whilst areas requiring further medium- to long-term research effort have been identified. Although potential has been realized to a very significant extent the technology is still at a formative stage. Extensive applied and fundamental research is required to ensure best utilization of available computing resources by improving the overall efficiency of the various techniques and to promote the emergence of new paradigms.
The potential of adaptive/evolutionary computing within the design field goes far beyond the optimization of specific complex systems. The development of appropriate co-operative strategies involving the various techniques and technologies can provide extensive support in multi-disciplinary design domains characterized by uncertainty relating to poor problem definition, multiple objectives and variable constraint. A breadth first adaptive exploratory approach can enhance innovative problem solving and provide decision support to establish optimal design direction. It is also possible that co-operative human/adaptive computing procedures can lead to the emergence of creative design solutions. This area of widening adaptive computing potential within engineering is now receiving significant attention from several international research groups and prototype co-operative search tools are gradually emerging. Complex issues relating to the integration of engineering heuristics with the adaptive processes and a human-centred interactive role for adaptive search processes within design team activities must also be considered.
A diverse set of papers is presented within this special issue. Engineering disciplines addressed include aerospace, automotive, electronics, mechanical, medical and communications. Several evolutionary approaches are included such as genetic algorithms, genetic programming, evolutionary programming and simulated annealing. Generic problem areas relating to multi-objective and constraint satisfaction are evident in some papers whereas several papers address the integration of heuristics in various forms to support the search and optimization process.
Fan Zhun and colleagues from the GARAGe group at Michigan State University introduce the integration of Bond Graphs with genetic programming to address mixed domain optimization problems relating to analog and MEM filter design. Masahiro Kanazaki and colleagues from the Institute of Fluid Dynamics, Tohoku University, Japan utilize a divided-range multi-objective genetic algorithm to optimize automotive exhaust manifold shapes with tapered pipes. A multi-objective simulated annealing algorithm provides the search capability for the discovery of radical bicycle frame designs for A. Suppapitnarm and G. Parks from the Engineering Design Centre at the University of Cambridge, UK. Abdullah Konak of Penn State University and Alice Smith from Auburn University, Alabama comprehensively tackle a complex capacitated network design problem treating cost, network performance and surviveability concurrently in an evolutionary multi-objective approach which also considers complex network constraints.
A particularly interesting problem relating to the optimization of minute, structurally complex cardio-vascular stents is addressed by Mark Atherton of South Bank University, UK and Ron Bates, now at the London School of Economics. Due to computationally expensive CFD analysis associated with the evaluation of stent performance, population-based stochastic optimization faces particular problems relating to the necessary small population sizes and minimal evaluation calls.
Carlos Coello-Coello and Ricardo Lanada-Becerra from the Evolutionary Computing Group at CINVESTAV-IPN, Mexico investigate the use of a cultural algorithm which supports the integration of experiential knowledge with an evolutionary programming optimization process. Sushil Louis from the University of Nevada takes a differing approach, augmenting genetic algorithm capabilities through the injection of intermediate design solutions taken from a case-based memory of past design problem solving. Chris Bonham and myself present a paper relating to recent developments of Cluster-oriented Genetic Algorithms (COGAs) which provide a design search and exploration capability particularly suited to the capture of experiential knowledge through interactive evolutionary design processes.
All-in-all this collection of papers appears to span the spectrum of evolutionary design capability from design search, exploration and discovery through to complex detailed design optimization. Many of the authors have an extensive history in the evolutionary computation field and the quality of the papers reflects their expertise and knowledge and the dedication of their researchers. I would like to thank all involved both for the papers submitted and for their efforts relating to the reviewing process.