Model-based Reinforcement Learning with Parametrized Physical Models and Optimism-Driven Exploration

摘要: In this paper, we present a robotic model-based reinforcement learning method that combines ideas from model identification and predictive control. We use feature-based representation of the dynamics allows to be fitted with simple least squares procedure, features are identified high-level specification robot's morphology, consisting number connectivity structure its links. Model control is then used choose actions under an optimistic dynamics, which produces efficient goal-directed exploration strategy. real time experimental results on standard benchmark problems involving pendulum, cartpole, double pendulum systems. Experiments indicate our able learn range tasks substantially faster than previous best methods. To evaluate approach realistic task, also demonstrate simulated 7 degree freedom arm.