HyperPPO is a reinforcement learning algorithm that uses graph hypernetworks to estimate the weights of multiple neural architectures at once. It is designed to find smaller, more efficient neural network architectures for controlling robots, particularly those with memory limitations. The algorithm allows users to select a network architecture that fits their computational constraints and performance needs. The study demonstrated the use of HyperPPO in the decentralized control of a Crazyflie2.1 quadrotor.
Publication date: 29 Sep 2023
Project Page: https://sites.google.com/usc.edu/hyperppo
Paper: https://arxiv.org/pdf/2309.16663
