Shingo Nakamura and Shuji Hashimoto
We describe the adaptive modeling of a physical system using the affine transform and its application to machine learning. We previously proposed a method to implement machine learning in physical hardware, where we built a simulator based on actual hardware input/output, and used it to optimize a controller. The method decreases stress on hardware because the controller is optimized by software via the simulator. Moreover, it does not require specific physical information on hardware. We also did not need to formulate hardware kinematics. When hardware changes, however, optimization must be redone to build the simulator -a clearly inefficient procedure. We therefore considered using previous optimization results when reoptimizing for new hardware. In the physical system, the aspect of the phase space does not vary much if the system structure remains the same. We applied affine transform to phase space of the physical system, to remodel the simulator for new hardware characteristics triggered by parameter changes. We used the remodeled simulator in machine learning to reoptimize the controller. In experiments, we used the swing-up pendulum problem to evaluate our proposal, comparing our proposal and original methods and finding that our proposal accelerates reoptimization.
Keywords: Simulator Building, Reinforcement Learning, Neural Network, Swing-Up Pendulum, Affine Transform