This repository contains the Matlab and Python implementation of R-FORCE (Robust FORCE)
We present a new initialization method for First Order Reduced and Controlled Error (FORCE) to achieve more robust performance to chaos. FORCE learning performs well only in a narrow range of chaos, but we want to explore whether there exists an alternative initialization method which can make FORCE more robust. In this paper, we demonstrate how we came up with idea of R-FORCE by treating RRNN as a dynamical system and describe how to generate R-FORCE specifically. Then we show the comparison results of R-FORCE, FORCE and Full-FORCE in target-learning and multi-dimension body modeling tasks. This results imply R-FORCE outperforms other state-of-art methods in terms of both mean absolute error and confidence interval.
Predicted Modeling Video
Training Instructions for One-dimension target funcion
- Run main.m for training, testing and plotting
Training Instruction for multi-dimension body modeling
- We used training data from UI-PRMD. Inside the data folder, we posted some of the original movement data (xxx.mat) and augmented data(xxxPeriod.mat).
- Run mainRFORCEMovementSimulation.m for training and testing. The testing simulation result and ground truth is saved as Simulation_skel and skel, respectively.
- We also included a trained model and simulation results for deep squat inside result.
- To visualize those results, you can use movementVisualization.m