欢迎您！东篱公司
退出
-
申报数据库
-
立项数据库
-
成果数据库
-
项目获奖数据库
中文摘要：
研究了几类典型增强学习算法的性能评估问题,包括Q-学习算法、最小二乘策略迭代（LSPI）和基于核的最小二乘策略迭代（KLSPI）算法等,重点针对Markov决策问题（MDP）的值函数平滑特性对算法性能的影响进行了研究。分别利用值函数非平滑的组合优化问题——旅行商问题（TSP）和值函数平滑的Mountain-Car运动控制问题,对不同增强学习算法的性能进行了测试和比较分析。分析了三种算法针对不同类型问题的各自特点,通过实验对比,验证了近似策略迭代算法,特别是KLSPI算法在解决值函数平滑的序贯决策问题时性能更优。通过分析实验结果表明,MDP值函数的平滑程度是影响近似策略迭代算法性能表现的重要因素。
英文摘要：
This paper studied the performance evaluation problem for reinforcement learning （ RL） algorithms,including Q-learning,least-squares policy iteration（ LSPI） and kernel based least-squares policy iteration（ KLSPI） . Investigated the performance influence of the smoothness of value functions in Markov decision processes in detail. Tested the RL algorithms on a combinatorial optimization problem—the traveling salesman problem （ TSP） ,which had non-smooth value functions and the Mountain-Car motion control problem with smooth value functions. Analyzed the characteristics of different RL algorithms and demonstrated that approximate policy iteration algorithms,especial KLSPI,had better performance when solving sequential decision-making problems with smooth value functions. Furthermore,it verifies that whether is the sequential decision-making problems with smooth value functions or not will play an important role in the performance of approximate policy iteration.
同期刊论文项目
同项目期刊论文
A spatial orthogonal allocation and heterogeneous cultural hybrid algorithm for multi-robot explorat
Vision-based long-distance lane perception and front vehicle location for full autonomous vehicles o
A regularity model-based multiobjective estimation of distribution algorithm with reducing redundant
Fault detection and identification for dead reckoning system of mobile robot based on fuzzy logic pa
Continuous-action reinforcement learning with fast policy search and adaptive basis function selecti
期刊信息
位置：