Reducing Evaluation Fatigue in Interactive Evolutionary Algorithms by Using an Incremental Learning Approach

摘要

Human fatigue is one of the most important problems of Interactive Evolutionary Algorithms (IEAs) that requires addressing. The problem of fatigue usually arises out of intensive interaction between the IEA system and the respondent. Consequently, due to lack of interest or attention, the respondent will provide biased answers either intentionally or unintentionally. To reduce the times of interaction in IEAs effectively, we adopt a learning approach to learn responedent's preference model and then use this model to predict the fitness values of any given individuals. Unlike other research, we propose a novel system called ALP-IGA that integrates the theorem of incremental machine learning, the Algorithmic Probability (ALP), with Interactive Genetic Algorithm (IGA). Since the ALP model will utilize each interaction effectively to improve the accuracy of prediction, it is very likely our ALP-IGA system can predict respondent's preferences precisely just after a few interactions. We have investigated the performance of our ALP-IGA via a Monte Carlo simulation. Experimental results indicated that the number of interactions needed by ALP-IGA is very small for some cases. In addition, we have also compared the prediction correctness of ALP-IGA with a contrast IEA system whose learning scheme is implemented by a neural network algorithm. The results showed that ALP-IGA is superior to IEA with neural network for all cases.