基于多层学习克隆选择的改进式增量型超限学习机算法

摘要

The great number of redundant nodes in an incremental extreme learning machine (I–ELM) may lower the learning efficiency of the algorithm, and complicate the network structure. To deal with this problem, we propose the improved I–ELM with kernel (I–ELMK) on the basis of multi-learning clonal selection algorithm (MLCSA). The MLCSA uses Baldwinian learning and Lamarckian learning, to exploit the search space by employing the information of antibodies, and reinforce the exploitation capacity of individual information. The proposed algorithm can limit the number of hidden layer neurons effectively to obtain more compact network architecture. The simulations show that MLCSI–ELMK has higher prediction accuracies online and off-line, while providing a better capacity of generalization compared with other algorithms.%针对增量型超限学习机(incremental extreme learning machine, I–ELM)中大量冗余节点可导致算法学习效率降低,网络结构复杂化等问题,提出基于多层学习(multi-learning)优化克隆选择算法(clone selection algorithm, CSA)的改进式I–ELM.利用Baldwinian learning操作改变抗体信息的搜索范围,结合Lamarckian learning操作提高CSA的搜索能力.改进后的算法能够有效控制I–ELM的隐含层节点数,使网络结构更加紧凑,提高算法精度.仿真结果表明,所提出的基于多层学习克隆选择的增量型核超限学习机(multi-learning clonal selection I–ELMK, MLCSI–ELMK)算法能够有效简化网络结构,并保持较好的泛化能力,较强的学习能力和在线预测能力.