CAJ | 학술논문

该文研究了量子理论与量子神经网络原理，深入分析了量子前向对传网模型与基于递归加权最小二乘的量子前向对传算法。提出了量子前向对传网的定义与知识集，提出了自适应量子前向对传算法，证明了算法的收敛性。该算法全面考虑了在本次学习之前学习速率的总体状况，通过自适应地改变学习速率，控制学习速率适时变化，改善网络的收敛性。有效克服了学习速率过高导致网络振荡发散与学习速率太小降低网络收敛速度的缺陷。仿真结果表明，自适应量子前向对传算法相对基于递归加权最小二乘的量子前向对传算法具有较少的网络训练迭代次数和较高的分类精度。
해문연구료양자이론여양자신경망락원리，심입분석료양자전향대전망모형여기우체귀가권최소이승적양자전향대전산법。제출료양자전향대전망적정의여지식집，제출료자괄응양자전향대전산법，증명료산법적수렴성。해산법전면고필료재본차학습지전학습속솔적총체상황，통과자괄응지개변학습속솔，공제학습속솔괄시변화，개선망락적수렴성。유효극복료학습속솔과고도치망락진탕발산여학습속솔태소강저망락수렴속도적결함。방진결과표명，자괄응양자전향대전산법상대기우체귀가권최소이승적양자전향대전산법구유교소적망락훈련질대차수화교고적분류정도。
This paper studies the quantum theory and the principle of Quantum Neural Network (QNN). Model of Quantum Forward Counter Propagation Neural Network (QFCPNN) and Recursive Weighted Least Squares Quantum Forward Counter Propagation Algorithm (RWLS_QFCPA) are analyzed. Definition and knowledge set of QFCPNN is proposed. Adaptive Quantum Forward Counter Propagation Algorithm (AQFCPA) is proposed and its convergence is proved. Full account of overall situations of learning rates before current learning, this algorithm improves network convergence by adaptively changing the learning rate and controls timely changing learning rate. This new algorithm effectively overcomes some defects including network oscillations divergence due to high learning rate and reducing network convergence speed due to low learning rate. The simulation results indicate that AQFCPA has less number of iterations of network training and higher classification accuracy relative to RWLS_QFCPA.