On the Synergy of Evolutionary Algorithms and Machine Learning
DescriptionIn recent years, evolutionary algorithms (EA) have become an increasingly influential meta-heuristic for solving timing consuming and expensive optimization problems in engineering. The idea of EA is to treat an unknown optimization problem as a black box, then use a nature-inspired algorithm to find a good solution to the problem quickly. Due to their general designs, they find promising applications in complex engineering problems with difficult, rugged landscapes, which often defy conventional optimization methods. The numerous success stories of applying EA, in both academia and in industry, testify that it is a powerful search paradigm. Some well known examples of EA are genetic algorithms, evolutionary strategies, swarm intelligence, differential evolution, artificial immune system, etc.On the other hand, machine learning (ML) has also made great strides of progress recently. Powerful learning methods have been developed to respond to the challenges in engineering applications such as artificial intelligence, data mining and robotics. The idea of ML is to reverse engineer the learning phenomena in animals and humans. The prediction quality in terms of generalization errors is an important yardstick in ML; looking from another angle, ML may also be regarded as tools that predict the space time problem landscape – the fitness landscape in EA terminology. Likewise, there are many successful applications of ML in academia and industry.Both EA and ML are powerful, appealing research paradigms. Surprisingly, there are relatively little studies on a synergy or integration of the two methods; and existing methods have some fundamental deficiencies. In this research, better methods for the synergy of EA and ML will be studied, leading to more powerful EA and ML hybrid meta-heuristics. The new method is an EA assisted by ML. Its performance will be verified on bench mark problems in the EA literature. The performance will then be studied on real engineering problems involving expensive and time consuming fitness evaluations. It is expected that the project will lead to novel EA methods capable of both evolution and learning, superior in performance to conventional EA methods. It will also lead to pioneering research thanks to the flow and interchange of ideas and perspectives from these two very active fields. Finally, the research may open up new practical applications for EA and ML.
|Effective start/end date||1/11/09 → 26/06/13|