Towards a common implementation of reinforcement learning for multiple robotic tasks
作者:
Highlights:
• A value-iteration-based algorithm is effectively applied to multiple robotic tasks.
• Novel, efficient method improves softmax action selection by structuring the inputs.
• Resulted RL method improves model-free solutions with minimal task knowledge.
• Compatible with high-dimensional RL with low computational cost.
• A new software framework allows us to test and compare diverse RL algorithms.
摘要
•A value-iteration-based algorithm is effectively applied to multiple robotic tasks.•Novel, efficient method improves softmax action selection by structuring the inputs.•Resulted RL method improves model-free solutions with minimal task knowledge.•Compatible with high-dimensional RL with low computational cost.•A new software framework allows us to test and compare diverse RL algorithms.
论文关键词:Reinforcement learning,Robotics,Exploration
论文评审过程:Received 5 April 2017, Revised 4 October 2017, Accepted 5 November 2017, Available online 6 November 2017, Version of Record 21 February 2018.
论文官网地址:https://doi.org/10.1016/j.eswa.2017.11.011