有限样本下模型选择理论与方法研究

Theory and method of model selection under limited samples

从理论上给出有限样本下模型选择的方法,详细讨论了样本噪声对模型选择的影响,提出了一种次优模型选择算法,并进一步给出了一种推挽式模型选择算法,它能够有效地提高模型选择的学习速度。两种算法具有很强的抗噪声能力,预测模型也具有很好的推广性。最后通过具体数值试验验证了上述理论和方法的可行性和优越性。

The method of model selection with limited samples is given in theory, and the noise's effect on the model selection is discussed. A suboptimal algorithm for model selection is proposed, then a push-pull algorithm for model selection is further presented, which can improve the learning speed of model selection effectively. The two algorithms have excellent anti-noise performance, and the prediction model has strong ability of generalization. The reliability and advantage of the above theory and method are illustrated through tests.