Generalised Task Planning with First-Order Function Approximation
Generalised Task Planning with First-Order Function Approximation
复制标题
使用一阶函数逼近的广义任务规划
DOI：
--
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Ronald P. A. Petrick
中科院分区：
文献类型：
--
作者：
Jun Hao Alvin Ng;Ronald P. A. Petrick
真实的世界机器人技术经常在不确定和动态的环境中运行。因此，对不同的规划问题进行概括是有实际意义的.在没有模型的情况下，基于值的强化学习可以用来学习目标导向的策略。通常，机器人与环境中的物体之间的交互呈现出一阶结构。我们建议使用5个一阶特征来近似生成广义策略的Q函数。服务机器人领域的实证结果表明，我们的在线7关系强化学习的方法是可扩展的大规模的问题，并使8跨不同的问题和仿真环境与DIS-9类似的过渡动力学迁移学习。10
: Real world robotics often operates in uncertain and dynamic environ-1 ments. Therefore, generalisation over different planning problems is of practical 2 interest. In the absence of a model, value-based reinforcement learning can be 3 used to learn a goal-directed policy. Typically, the interaction between robots and 4 the objects in the environment exhibit a first-order structure. We propose to use 5 first-order features to approximate the Q-function that generates a generalised pol-6 icy. Empirical results for a service robot domain show that our method of online 7 relational reinforcement learning is scalable to large-scale problems and enables 8 transfer learning across different problems and simulation environments with dis-9 similar transition dynamics. 10