摘要
本文提出了一种新的基于Agent的神经网络隐层结构的优化算法(OHA)。该方法包括两个部分,分别由RLAgent和NNAgent合作完成。RLAgent根据强化学习算法找到一个比当前节点数更优的解,并反馈给NNAgent。NNAgent据此构建相应的网络,并采用分层训练的算法对该网络进行优化,训练结果再发给RLAgent。在多次循环后,OHA算法就可以找到一个训练误差最小的全局最优解(权值及隐层节点数)。本文讨论了有关的算法、测试和结果分析。Iris数据集和危险评估数据集的测试结果表明,算法避免了盲目搜索造成的计算开销,明显改善了优化性能。
This paper proposes a novel algorithm to Optimize the number of Hidden nodes based on Agent(OHA). This approach is completed by two cooperating agents, the RL agent and the NN agent. The RL agent searches better number of hidden nodes according to the reinforcement learning method, and the NN agent optimizes the weights of network with the number by using the separate learning algorithm. After much running, the best solution(weights and hidden nodes) is loca- ted. The optimization algorithms and tests are discussed. The test results obtained by using the Iris data set and the risk e- valuation data set show the algorithm is better than those by the most commonly used optimization techniaues.
出处
《计算机工程与科学》
CSCD
北大核心
2010年第5期30-33,共4页
Computer Engineering & Science
关键词
神经网络
隐层节点
隐层结构优化
智能代理
强化学习
neural networks
hidden node
hidden-layer architecture optimization
agent
reinforcement learning