Abstract: This paper improves the authors' previously proposed nonlinear active hand- writing models and applies them into radical extraction for handwritten Chinese character recognition. In the training phase, kernel principal component analysis is used to capture nonlinear handwriting variations. Then deformable models can be generated by varying a small number of shape parameters. In the recognition phase, genetic algorithms, rather than dynamic tunneling algorithm in the original version, are employed to search for the optimal shape parameters. Experiments are conducted on 200 radicals covering 2154 character categories. The correct matching rate is 97.4% on 430,800 loosely-constrained characters. Comparison with existing representative radical approaches shows that our method achieves superior performance.