【作者】 肖伟；

【导师】 周东辉；

【作者基本信息】 中国海洋大学 ， 通信与信息系统， 2005， 硕士

【摘要】 本文在采用BP学习算法的多层感知器的基础上,考虑初始权值的波动和样本组的因素,提出了初始权值优化技术,以提高初始权值与样本组的匹配能力。取得以下的结论: 在采用BP学习算法的多层感知器的情况下,采用单输出型方式,初始权值优化技术在有限的前几次优化中,取得较大的峰值的概率较大;而采用多输出型方式,并且样本的维数与神经元数目增加的情况下,初始权值优化技术在有限的几次优化中,取得较大的峰值的概率很小。因此,初始权值优化技术适合于单输出型方式。 本文还对初始权值优化技术的优化过程进行了分析,得出以下结论: 权值矩阵的F—范数在优化函数值变化较大时并无很大的波动;而对不同优化函数值的权值矩阵,对它们的差构成的新矩阵,通过研究新矩阵F—范数的研究发现,不同优化函数值对应的权值矩阵的差别是很明显的。初始权值优化技术选择了更适合当前样本组的初始权值组。 对在多输出型情况下,优化函数不容易取得较大峰值的原因: 由于多输出型要以一个网络来适应样本组的所有样本,在训练过程中,中间层权值的更新存在相互影响的现象;而在单输出型中,使用的是网络群结构,克服了类别之间的耦合,中间层权值更新不存在相互影响的现象。优化函数的自变量数目在单输出型的情况下比多输出型要多几倍,并且自变量之间不存在相互影响的问题,因此,寻优的过程变得容易;而在多输出型情况下,自变量在权值更新过程中的相互影响,优化函出现较大的峰值的概率降低。 本文的室内移动机器人采用的是单输出型方式,结合SPCE061A对语音的软硬件支持,通过语音教学的方式,由传感器组和语音编码后的向量构成样本,以语音触发的方式让机器人在未知的环境中采样,采集完成后,机器人通过采集的样本组来训练自己。训练完成后,通过学习训练后的网络直接控制移动机器人的动作。论文对采用初始权值优化技术的和未采用的网络分别作了相应的实验,实验结果表明,初始权值优化技术显著地提高了收敛的速度。 由初始权值优化技术的启发,本文对情感思维模型作了相关的探讨,并且对情感和思维的协调提出了相应的解决方案,并编写了测试软件,对加载选择器的合理性进行测试,得出结论:这种解决方案对协调情感和思维是合理的。更多还原

【Abstract】 Based on multiplayer feed-forward neural networks using BP algorithm, considering the fluctuation of initial weights and sample sets, a novel optimal weights initialization technology is proposed for the sake of matching between initial weights and sample sets. In this paper several conclusions are gained:In the condition of adopting multiplayer feed-forward neural network with one output (called single output model below) using BP algorithm, optimal weights initialization technology can gain bigger peak value within several times with great possibility. In the condition of adopting neural network with multiple outputs (called multiple output model below), and in increasing sample dimension and neural network neurons, optimal weights initialization technology can not achieve better initial weights group within several finite steps. As a result, optimal weights initialization technology is suitable to single output model.Analysis on the process of optimal weights initialization technology is achieved and the conclusions made are as follows:The F-norm of weights matrix has little fluctuation when the value of optimal function changes greatly. As to matrixes of different optimal function values, a new matrix, which is the balance between two weights matrix with different optimal function values, is constructed. Analysis on the value of new matrix’ s F-norm shows that the difference between two weights matrixes is evident. Optimal weights initialization technology chooses the better weights group for current sample sets.As to the poor performance in achieving big peak value using neural networks with multiple output, a research is made and the conclusions are as follows:To multiple output model, a single network must adapt to all sample sets, and there exists weights’ influence in the training process; However in single output model, network group structure is adopted, which conquers the interaction between categories as well as in the weights’ change process. To the single output model, the number of variables is several times to multiple output model, and there is no influence between variables. Consequently, the optimal process becomes effortless; To the multiple output model, because of the influence in weights’ change process, the appearance possibility of peak value of optimal function is reduced greatly.In this paper, the mobile robot adopts single output model. With the SPCE061A’ s voice features, samples are constituted by the encoded output of sensor group and voice teaching. And samples are gained by voice triggering. Robot trains itself with the sample group gathered during samples collection. After the training process, the mobile robot is controlled by the trained neural network directly. In order to observe the difference between the network which adopts optimal weights initialization technology and the one does not, an experiment is done. The result shows that optimal weights initialization technology enhances the speed of convergence greatly.With the inspiration of optimal weights initialization technology, a research on emotion and thinking model is discussed and a solution to balance emotion and thinking is proposed. In order to test the rationality of the loading machine, test software is accomplished. The result shows that the solution to emotion and thinking is rational.更多还原