eduzhai > Applied Sciences > Engineering >

Reinforcement Learning Control of Robotic Knee with Human in the Loop by Flexible Policy Iteration

  • Save

... pages left unread,continue reading

Document pages: 14 pages

Abstract: We are motivated by the real challenges presented in a human-robot system todevelop new designs that are efficient at data level and with performanceguarantees such as stability and optimality at systems level. Existingapproximate adaptive dynamic programming (ADP) results that consider systemperformance theoretically are not readily providing practically useful learningcontrol algorithms for this problem; and reinforcement learning (RL) algorithmsthat address the issue of data efficiency usually do not have performanceguarantees for the controlled system. This study fills these important voids byintroducing innovative features to the policy iteration algorithm. We introduceflexible policy iteration (FPI), which can flexibly and organically integrateexperience replay and supplemental values from prior experience into the RLcontroller. We show system level performances including convergence of theapproximate value function, (sub)optimality of the solution, and stability ofthe system. We demonstrate the effectiveness of the FPI via realisticsimulations of the human-robot system. It is noted that the problem we face inthis study may be difficult to address by design methods based on classicalcontrol theory as it is nearly impossible to obtain a customized mathematicalmodel of a human-robot system either online or offline. The results we haveobtained also indicate the great potential of RL control to solving realisticand challenging problems with high dimensional control inputs.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...