首页|Studies from University of Alberta in the Area of Robotics Described (Explainabi lity of Deep Reinforcement Learning Algorithms In Robotic Domains By Using Layer -wise Relevance Propagation)
Studies from University of Alberta in the Area of Robotics Described (Explainabi lity of Deep Reinforcement Learning Algorithms In Robotic Domains By Using Layer -wise Relevance Propagation)
扫码查看
点击上方二维码区域,可以放大扫码查看
原文链接
NETL
NSTL
By a News Reporter-Staff News Editor at Robotics & Machine Learning DailyNews-Data detailed on Robotics have been presented. Ac cording to news reporting out of Edmonton,Canada, by NewsRx editors, research s tated, "A key component to the recent success of reinforcementlearning is the i ntroduction of neural networks for representation learning. Doing so allows for solvingchallenging problems in several domains, one of which is robotics."Funders for this research include Mitsubishi, Alberta Machine Intelligence Insti tute. Our news journalists obtained a quote from the research from the University of A lberta, "However,a major criticism of deep reinforcement learning (DRL) algorit hms is their lack of explainability andinterpretability. This problem is even e xacerbated in robotics as they oftentimes cohabitate space withhumans, making i t imperative to be able to reason about their behavior. In this paper, we propos e toanalyze the learned representation in a robotic setting by utilizing Graph Networks (GNs). Using the GNand Layer-wise Relevance Propagation (LRP), we repr esent the observations as an entity-relationship toallow us to interpret the le arned policy. We evaluate our approach in two environments in MuJoCo. Thesetwo environments were delicately designed to effectively measure the value of knowle dge gained by ourapproach to analyzing learned representations. This approach a llows us to analyze not only how differentparts of the observation space contri bute to the decision-making process but also differentiate betweenpolicies and their differences in performance. This difference in performance also allows for reasoning aboutthe agent's recovery from faults."
EdmontonCanadaNorth and Central Amer icaAlgorithmsEmerging TechnologiesMachine LearningReinforcement LearningRoboticsRobotsUniversity of Alberta