Proceedings of the Fourth International Conference on Machine Learning and Cybernetics, Guangzhou, 18-21 August 2005 AN IMPROVED INTERACTIVE GENETIC ALGORITHM INCORPORATING RELEVANT FEEDBACK SHANG-FEI WANG, XU-FA WANG, JIA XUE Department of Computer Science and Technology, University of Science and Technology of China, Anhui, 230027, China E-MAIL: sfwang@ustc.edu.cn Abstract: This paper has proposed a new Interactive Genetic Algorithm (IGA) framework Incorporating Relevant Feedback (RF), in which human evaluation is regarded as not only the fitness function of GA, but also the relevant score to instruct interactive machine learning. Thus, on the one hand, user's fatigue, the key issue of IGA, can be alleviated, since some individuals with higher preference weight are added in each generation through relevance feedback technology. On the other hand, the two mapping functions between the low-level parameter space and the high-level users' psychological space can be built during interactions. An instance of this frame, which uses Support Vector Machine (SVM) as the machine learning method in RF, is also provided. The effectiveness of our approach is first evaluated through simulation tests using two benchmark functions. The experimental results show that the convergence speed of the proposal is much faster than that of normal IGA. Then, the approach is applied to retrieve images with emotion semantics queries. The subject experiments also demonstrate that the proposal algorithm can alleviate user fatigue. Furthermore, SVM constructs an individual emotion user model though learning. Keywords: Interactive genetic algorithm; relevant feedback; support vector machines; emotion semantics; image retrieval 1. Introduction Digital image retrieval systems allow sophisticated querying and searching by image content. Since 1990's, Content-Based Image Retrieval (CBIR) has attracted great research attention [1]. Early research focused on finding the best representation for image features. The similarity between two images is calculated by summing the distances of low-level features with fixed weights. In this context, high-level concepts and user's subjectivity cannot be well modeled. Recent approaches introduce human-computer interaction into CBIR. One approach is Interactive Genetic Algorithm (IGA) [2], which is an optimization technology that adopts GA for system optimization based on human’s subjective evaluation. Human observe system outputs and give evaluations based on the similarity between the system outputs and their goals in the high-level semantic or concept space, and GA searches in the low-level feature space. Using this technology, the high-level concept and subjectivity borne users can be automatically captured by the system to some degree. However, human fatigue is one of the biggest problems to any human-computer interaction approach, especially to IGA [3], since users should evaluates individuals in each generation. Another approach is Relevance Feedback (RF), in which users can select the most relevant images and provide a weight of preference for each relevant image to refine the query [4]. Various RF algorithms from heuristic-based feature weighting schemes to optimal learning algorithms have been proposed in the area of CBIR, which have been a proper technology to build the mapping function from the low-level features of images to the high-level of human semantics. However, RF also has its difficulties, such as small sample issue. This paper provides a new framework, which incorporating Interactive Genetic Algorithm with Relevance Feedback (IGARF). In initial stage of IGA, the user's satisfied individuals are few. These few good individuals are reserved to reproduce the next generation, while most individuals with low fitness are eliminated. However, these bad individuals also indicate the user's negative impression. In our proposed IGARF, those good individuals and bad ones are respectively regarded as positive and negative examples, which are used to train a classification on-line. Then all images in the database are classified into two classes: positive and negative class. Some best images in the positive class with higher preference weight are added to each generation of IGA. Thus the number of good individuals in each generation increases, and the acceleration of GA convergence is expected. Meanwhile the system also constructs two mapping functions between parameter space to psychological space through GA and RF. Support Vector Machine (SVM) is applied as the machine learning method since it is a small sample model. We first 0-7803-9091-1/05/$20.00 ©2005 IEEE 2996