Two main practical problems arise when dealing with autonomous learning of the control of Linked Multi-Component Robotic Systems (L-MCRS) with Reinforcement Learning (RL): time and space consumption, due to the convergence conditions of the RL algorithm applied, i.e. Q-Learning algorithm, and the complexity of the system model. Model approximate response allows to speedup the realization of RL experiments. We have used a multivariate regression approximation model based on Artificial Neural Networks (ANN), which has achieved a 90% and 27% of time and space savings compared to the conventional Geometrically Exact Dynamic Splines (GEDS) model.
展开▼
