Evolutionary Complex Engineering Optimization: Opportunities and Challenges

Digital Object Identifier 10.1109/MCI.2013.2264563 Evolutionary algor ithms (EAs) including other meta-heuristics such as particle swarm optimization and differential evolution have shown to be powerful for global optimization of a wide range of problems. In recent years, huge research effort has been devoted to solving complex engineering optimization (CEO) problems. Among others, CEO problems are often subject to large amount of uncertainties, such as varying environmental conditions, system degeneration, or changing customer demand; they are highly constrained, where the constraints themselves can also change over time; computationally expensive and need to satisfy multiple criteria involving multiple disciplines; they usually consist of mutually dependent subsystems having a high number of possibly correlated decision variables. Furthermore, complex engineering optimization in most cases is embedded into a larger design process involving several teams and tools working sequentially and in parallel on a variety of temporally and spatially decomposed sub-systems. As a result, a number of new research areas have emerged, including evolutionary optimization in dynamic and uncertain environments [1], [2], surrogate-assisted evolutionary optimization [3], multiand many-objective optimization [4], [5], large-scale optimization [6], and integrated control and optimization [7], [8], just to name a few. Despite the fact that the above topics are motivated from real-world challenges, not much of the research results in the above areas have been applied to solving real-world problems and most challenges in complex engineering optimization remain unsolved. Thus, concerns have been raised about the relevance of these lines of research to real-world problems. First, it remains unclear whether the challenges addressed in these areas are of practical significance in the real world. For example, in evolutionary dynamic optimization, most algorithms have been designed to closely track the changing optimum. This is ideal in principle if the designed algorithm is able to follow the moving optimum at any time instant. However, frequent changes of the optimal design will not only be constrained by time for implementing the new designs, but will also incur very high cost, which makes it impractical in the realworld. One recent idea to address this issue is to find optimal solutions that are robust over time so that optimal solutions that change most slowly will be identified to minimize the need to change the design [9], which can be seen as a trade-off between optimum tracking and robustness. Second, a large number of test problems have been designed for benchmarking the performance of different metaheuristics. Such test problems are meant to reflect the hardness of real-world problems and have widely been used in dynamic optimization, multi-objective optimization, constrained optimization as well as in large-scale optimization. For publishing a paper, these test problems have become almost a standard for demonstrating the advantage of newly proposed algorithms over the state-of-the-art. However, little thought has been given to whether these popular suits of test problems are of significance in the real-world, i.e. how much the test problems are relevant to real optimization cases. Third, it is no longer straightforward to come up with one single well-defined performance indicator to compare the performance of the developed algorithms. This is true for multi-objective optimization, where the quality of the achieved solution set must be assessed using more than one performance indicator, including accuracy and diversity. In addition, the quality of solutions can also be subjective, often depending on the preference of a human decision maker. The situation becomes worse in manyobjective optimization, where the number of objectives considered is often very high. Obviously, comparing an extremely small set of solutions in a huge space makes little sense and can even become misleading without a clear preference. Even the seemingly straightforward visualization of the results of a many objective optimization process can be very awkward due to the high dimensionality of the objective space. To address the above-mentioned concerns, the first question one might raise is: What makes a CEO problem really difficult to solve? It is not straightforward to provide a simple answer to this question, as it is inherently problem dependent. In the following, we attempt to discuss a few points, which—we hope—can shed some light on the question. If one has had experience in solving real-world CEO problems, one will be Evolutionary Complex Engineering Optimization: Opportunities and Challenges