Mental workload(MW)-based adaptive system has been found to be an effective approach to enhance the performance of human-machine interaction and to avoid human error caused by overload.However,MW recognized from the spontaneously generated electroencephalogram(EEG)was found to be task-specific.In existing studies,EEG-based MW classifier can work well under the task used to train the classifier(within-task)but crash completely when used to classify MW of a task that is similar to but not included in the training data(cross-task).The possible causes have been considered to be the task-specific EEG patterns,the mismatched workload across tasks and the temporal effects.In this study,cross-task performance-based feature selection and regression model were tried to cope with these challenges,in order to make EEG-based MW estimator trained on working memory tasks work well under a complex simulated multi-attribute task(MAT).The results show that the performance of regression model trained on working memory task and tested on multi-attribute task with the feature subset picked-out were significantly improved(COR: 0.740±0.147 and 0.598±0.161 for feature selection data and validation data respectively)when compared to the performance in the same condition with all features(chance level).It can be inferred that there do exist some MW-related EEG features can be picked out and there are something in common between MW of a relatively simple task and a complex task.This study provides a promising approach to measure MW across tasks.