摘要
常规Ant-Q算法计算复杂度随问题的规模呈现出阶乘级的增长,极大地抑制了算法的收敛速度,同时其仅关注单一任务本身,使得求出的解不具有可重用性,在处理一系列相关联任务时效率较低.为此,提出一种基于知识迁移的Ant-Q算法,通过贝叶斯理论分析源任务与目标任务的相似率,并以此为权值确定各源任务的迁移样本数,然后将各源任务样本按迁移价值降序排列,筛选出有效迁移样本,指导Agent快速做出合理决策.在att532旅行商问题上的仿真结果表明,知识迁移能够有效降低目标任务的学习难度,从而快速找到问题的最优解.
The computational complexity of traditional Ant-Q algorithm shows factorial growth with the scale of the studied problem,which greatly reduces the convergence speed.Moreover,the traditional Ant-Q algorithm only focuses on a single task,therefore,the solution for the task cannot be reusable and the algorithm will handle a series of related tasks with low efficiency.In order to improve the convergence speed,a kind of Ant-Q algorithm based on knowledge transfer is proposed.At first,the similarity between each source task and a target task is computed according to the Bayesian theory.Then the obtained similarities are viewed as the weights to determine the number of samples transferred from every source task.In the third step,the samples from source tasks are listed in a descending order according to its transfer values and some valid samples are selected.In this way,the selected samples can guide an Agent to make a rational decision quickly.Simulation results involving a traveling salesman problem att532 illustrate that the knowledge transfer technology can effectively reduce the difficulty of learning a new task and quickly find an optimal solution.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2011年第10期2359-2365,共7页
Acta Electronica Sinica
基金
国家自然科学基金(No.60804022
No.60974050
No.61072094)
教育部新世纪优秀人才支持计划(No.NCET-08-0836
No.NCET-10-0765)
霍英东教育基金会青年教师基金(No.121066)
江苏省自然科学基金(No.BK2008126)
关键词
知识迁移
Ant-Q算法
贝叶斯理论
样本筛选
旅行商问题
knowledge transfer
ant-Q algorithm
bayesian theory
sample selection
traveling salesman problem
