Genetic Programming: An emerging engineering tool

It is now 15 years since Genetic Programming (GP) burst on the research scene in Koza’s seminal book [3], though it had been pre-figured by a number of previous explorations of similar ideas [1,7]. Koza first showed that GP could be a credible engineering tool, and subsequent volumes [4–6] and the human-competitive research competitions [2] have drawn attention to its wide range of achievements. The papers in this issue were submitted in response to the Asia-Pacific Symposium on Genetic Programming (www.aspgp.org), which aims to bring together researchers and practitioners in Genetic Programming from the Asia-Pacific region and internationally, and promote research in and application of GP in the region. The papers both confirm the wide applicability of GP, and illustrate the continuing ferment in its development. Two papers are directly concerned with real-world applications, one uses GP to improve a general engineering method, while the final two emphasise algorithm improvements aimed at improving the applicability and scalability of GP. Polimon et al. use GP to evolve block ciphers for cryptographic applications. They demonstrate the automated evolution of powerful ciphers. Their work demonstrates the strength of the ciphers that GP can generate, suggesting that GP could potentially generate competitive state-of-the-art ciphers. Equally important, the results suggest that GP can readily generate new, high-quality ciphers, inspiring a scenario in which ciphers might be application-specific, reducing the payoff for cracking any particular cipher, thereby reducing the incentive to undertake cracking at all. Beyond the particular problem under consideration, the paper demonstrates the ability of GP to generate results comparable with those generated by competent human practitioners. Andreae et al. use GP for automated stress detection in spoken English. The stress detection system is intended as a component of an automated learning support system for spoken English, aimed at detecting stress errors in the student’s speech production, and providing feedback. However it potentially has much wider application, as stress detection is important not only for detecting errors, but also for interpreting the meta-linguistic content of speech utterances in natural language understanding. In this task, GP evolved rules which detected stress significantly more reliably than either decision trees (DT) or support vector machines (SVM), presumably because they were able to automatically make use of interactions between attributes, which were beyond the expressive reach of DT or SVM representation languages. Pujol and Poli use GP to transform the decision space for parameter optimisation and tuning (POT) problems. Compared with standard methods, on a wide range of test sets, the GP-based approach yields better results, is more robust to perturbation of the test functions (though at cost of slower convergence), and more scalable than current methods. While this research is in its early stages, it appears that the use of GP to transform the space may represent a significant improvement in POT methods.